Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

test(normalize): enforce precision and hosted regression coverage

zhifu gao1ad7430b925e
diff --git a/.github/workflows/test-funasr-onnx-package.yml b/.github/workflows/test-funasr-onnx-package.yml
index 1a0e1438f..f94b86e09 100644
--- a/.github/workflows/test-funasr-onnx-package.yml
+++ b/.github/workflows/test-funasr-onnx-package.yml
@@ -13,6 +13,7 @@ on:
- "tests/test_funasr_onnx_paraformer_online_empty_final.py"
- "tests/test_funasr_onnx_vad_online_short_chunks.py"
- "tests/test_funasr_onnx_vad_multiple_inputs.py"
+ - "tests/test_funasr_onnx_punctuation_empty_input.py"
- "tests/test_rich_transcription_postprocess.py"
- "tests/test_rich_transcription_event_segments.py"
push:
@@ -29,6 +30,7 @@ on:
- "tests/test_funasr_onnx_paraformer_online_empty_final.py"
- "tests/test_funasr_onnx_vad_online_short_chunks.py"
- "tests/test_funasr_onnx_vad_multiple_inputs.py"
+ - "tests/test_funasr_onnx_punctuation_empty_input.py"
- "tests/test_rich_transcription_postprocess.py"
- "tests/test_rich_transcription_event_segments.py"
workflow_dispatch:
@@ -84,7 +86,7 @@ jobs:
import importlib.util
from importlib.metadata import requires, version
- assert version("funasr-onnx") == "0.4.7"
+ assert version("funasr-onnx") == "0.4.9"
assert importlib.util.find_spec("torch") is None
requirements = requires("funasr-onnx") or []
@@ -115,6 +117,9 @@ jobs:
- name: Verify offline VAD results per input
run: python tests/test_funasr_onnx_vad_multiple_inputs.py
+ - name: Verify empty punctuation streaming cache
+ run: python tests/test_funasr_onnx_punctuation_empty_input.py
+
- name: Verify source distribution error paths
run: |
python -m pip install --force-reinstall --no-deps dist/funasr-onnx/*.tar.gz
@@ -125,7 +130,7 @@ jobs:
if: matrix.python-version == '3.12'
uses: actions/upload-artifact@v7
with:
- name: funasr-onnx-0.4.7
+ name: funasr-onnx-0.4.9
path: dist/funasr-onnx/*
if-no-files-found: error
diff --git a/.github/workflows/test-kws-output.yml b/.github/workflows/test-kws-output.yml
index 55b2f4cb0..754e333e0 100644
--- a/.github/workflows/test-kws-output.yml
+++ b/.github/workflows/test-kws-output.yml
@@ -8,12 +8,14 @@ on:
- "funasr/models/sanm_kws/**"
- "funasr/models/sanm_kws_streaming/**"
- "funasr/frontends/wav_frontend.py"
+ - "funasr/models/fsmn_vad_streaming/model.py"
- "funasr/auto/auto_model.py"
- "funasr/utils/datadir_writer.py"
- "funasr/utils/kws_utils.py"
- "tests/test_kws_optional_output.py"
- "tests/test_kws_streaming_continuity.py"
- "tests/test_kws_utils.py"
+ - "tests/test_wav_frontend_no_lfr.py"
- "tests/test_fsmn_vad_streaming_buffers.py"
- "tests/test_dynamic_streaming_vad.py"
- "tests/test_fsmn_vad_dynamic_silence.py"
@@ -26,12 +28,14 @@ on:
- "funasr/models/sanm_kws/**"
- "funasr/models/sanm_kws_streaming/**"
- "funasr/frontends/wav_frontend.py"
+ - "funasr/models/fsmn_vad_streaming/model.py"
- "funasr/auto/auto_model.py"
- "funasr/utils/datadir_writer.py"
- "funasr/utils/kws_utils.py"
- "tests/test_kws_optional_output.py"
- "tests/test_kws_streaming_continuity.py"
- "tests/test_kws_utils.py"
+ - "tests/test_wav_frontend_no_lfr.py"
- "tests/test_fsmn_vad_streaming_buffers.py"
- "tests/test_dynamic_streaming_vad.py"
- "tests/test_fsmn_vad_dynamic_silence.py"
@@ -63,6 +67,7 @@ jobs:
tests/test_kws_optional_output.py \
tests/test_kws_streaming_continuity.py \
tests/test_kws_utils.py \
+ tests/test_wav_frontend_no_lfr.py \
tests/test_fsmn_vad_streaming_buffers.py \
tests/test_dynamic_streaming_vad.py \
tests/test_fsmn_vad_dynamic_silence.py
diff --git a/.github/workflows/test-numpy-compatibility.yml b/.github/workflows/test-numpy-compatibility.yml
index 288f6926c..a6f2b474c 100644
--- a/.github/workflows/test-numpy-compatibility.yml
+++ b/.github/workflows/test-numpy-compatibility.yml
@@ -8,6 +8,8 @@ on:
- "funasr/cli.py"
- "tests/test_cli.py"
- "funasr/auto/auto_model.py"
+ - "funasr/tokenizer/sentencepiece_tokenizer.py"
+ - "tests/test_sentencepiece_tokenizer.py"
- "tests/test_speaker_optional_dependency.py"
- "funasr/utils/load_utils.py"
- "funasr/bin/realtime_ws.py"
@@ -26,6 +28,8 @@ on:
- "tests/test_pcm_input_format.py"
- "tests/test_load_audio_bytes.py"
- "funasr/frontends/**"
+ - "funasr/models/normalize/utterance_mvn.py"
+ - "tests/test_utterance_mvn.py"
- "funasr/models/emotion2vec/**"
- "funasr/models/campplus/**"
- "funasr/models/sense_voice/model.py"
@@ -45,6 +49,8 @@ on:
- "funasr/cli.py"
- "tests/test_cli.py"
- "funasr/auto/auto_model.py"
+ - "funasr/tokenizer/sentencepiece_tokenizer.py"
+ - "tests/test_sentencepiece_tokenizer.py"
- "tests/test_speaker_optional_dependency.py"
- "funasr/utils/load_utils.py"
- "funasr/bin/realtime_ws.py"
@@ -63,6 +69,8 @@ on:
- "tests/test_pcm_input_format.py"
- "tests/test_load_audio_bytes.py"
- "funasr/frontends/**"
+ - "funasr/models/normalize/utterance_mvn.py"
+ - "tests/test_utterance_mvn.py"
- "funasr/models/emotion2vec/**"
- "funasr/models/campplus/**"
- "funasr/models/sense_voice/model.py"
@@ -119,12 +127,14 @@ jobs:
run: |
python -m pytest -q \
tests/test_numpy_compatibility.py \
+ tests/test_utterance_mvn.py \
tests/test_numpy_cluster_compatibility.py \
tests/test_cluster_backend.py \
tests/test_speaker_optional_dependency.py \
tests/test_eend_feature_compatibility.py \
tests/test_pypi_metadata.py \
tests/test_sensevoice_word_timestamps.py \
+ tests/test_sentencepiece_tokenizer.py \
tests/test_pcm_input_format.py \
tests/test_realtime_ws_service.py \
tests/test_realtime_ws_hf_overrides.py \
diff --git a/README.md b/README.md
index 277870f0a..f36b7a086 100644
--- a/README.md
+++ b/README.md
@@ -364,7 +364,9 @@ recordings with the same evaluation scope.
## What's new
- **MOSS-Transcribe-Diarize** brings long-form ASR, timestamps, and anonymous speaker labels to FunASR services, Docker, Kubernetes, vLLM/SGLang workflows, and FunClip. [Deploy MOSS ->](./docs/moss_transcribe_diarize.md)
-- **FunASR 1.4.21** preserves the streaming punctuation cache on empty or whitespace chunks, returning an empty result without touching cached words or the embedding path; nonempty inference is unchanged. [Fix #3824](https://github.com/modelscope/FunASR/pull/3824). Install with `python -m pip install -U "funasr==1.4.21"`. [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
+- **FunASR 1.4.23** fixes the online frontend cursor for `lfr_m=lfr_n=1` and redundant empty final offline VAD chunks. With a supplied `token_list`, SentencePiece `ids2tokens` returns model pieces; use `tokens2text` for text. Without `token_list`, legacy text decoding is unchanged. [Fix #3849](https://github.com/modelscope/FunASR/pull/3849), [fix #3783](https://github.com/modelscope/FunASR/pull/3783). Install with `python -m pip install -U "funasr==1.4.23"`. [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.23)
+ **FunASR 1.4.22** honors per-call `merge_length_s` and evaluates dynamic VAD silence thresholds per frame using the current segment duration, preserving the fixed-threshold opt-out. [Fix #3841](https://github.com/modelscope/FunASR/pull/3841), [fix #3833](https://github.com/modelscope/FunASR/pull/3833). [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.22)
+ **FunASR 1.4.21** preserves the streaming punctuation cache on empty or whitespace chunks; nonempty inference is unchanged. [Fix #3824](https://github.com/modelscope/FunASR/pull/3824). [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
**FunASR 1.4.20** preserves the original optional speaker dependency failure in a chained `ImportError` when `spk_model` is requested, while ASR-only construction remains available. [Fix #3819](https://github.com/modelscope/FunASR/pull/3819). [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.20)
**FunASR 1.4.19** handles empty-token Paraformer batches, enforces `max_speakers` for automatic clustering, and adds validated JSON-object `--hf_overrides` support to realtime WebSocket serving. [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.19)
Previous release **FunASR 1.4.18** fixes cross-script hotword misreplacement, filtered-token timestamp alignment, and Chinese VAD sentence/speaker drift. [Release and verification scope ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.18)
diff --git a/README_ja.md b/README_ja.md
index 13450d864..25ee7e109 100644
--- a/README_ja.md
+++ b/README_ja.md
@@ -112,7 +112,9 @@ CER/WER をそろえて比較してください。オフラインのスループ
## 最新情報
- **MOSS-Transcribe-Diarize** を FunASR service、Docker、Kubernetes、vLLM/SGLang workflow、FunClip に統合し、長時間 ASR、timestamp、匿名 speaker label を一度に処理できます。[MOSS をデプロイ ->](./docs/moss_transcribe_diarize.md)
-- **FunASR 1.4.21** はストリーミング句読点処理の空または空白入力で cache を保持し、キャッシュ済み単語や embedding 経路に触れず空結果を返します。非空推論は変更されません。[修正 #3824](https://github.com/modelscope/FunASR/pull/3824)。`python -m pip install -U "funasr==1.4.21"` でインストールできます。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
+- **FunASR 1.4.23** は `lfr_m=lfr_n=1` のオンライン特徴抽出のフレーム位置と、オフライン VAD の余分な空の最終チャンクを修正します。`token_list` 指定時の SentencePiece `ids2tokens` はモデルのピースを返し、テキストへの変換には `tokens2text` を使います。未指定時の従来のテキスト復号は維持します。[修正 #3849](https://github.com/modelscope/FunASR/pull/3849)、[修正 #3783](https://github.com/modelscope/FunASR/pull/3783)。`python -m pip install -U "funasr==1.4.23"` でインストールできます。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.23)
+ **FunASR 1.4.22** は呼び出しごとの `merge_length_s` を反映し、現在の発話区間の長さに基づいて動的 VAD 無音しきい値をフレーム単位で更新します。固定しきい値モードも維持します。[修正 #3841](https://github.com/modelscope/FunASR/pull/3841)、[修正 #3833](https://github.com/modelscope/FunASR/pull/3833)。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.22)
+ **FunASR 1.4.21** はストリーミング句読点処理の空入力で cache を保持します。非空推論は変更されません。[修正 #3824](https://github.com/modelscope/FunASR/pull/3824)。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
**FunASR 1.4.20** は `spk_model` 指定時に任意の話者依存モジュールを読み込めない場合、元の例外を連鎖した `ImportError` で報告します。ASR のみの構築は引き続き利用できます。[修正 #3819](https://github.com/modelscope/FunASR/pull/3819)。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.20)
**FunASR 1.4.19** は Paraformer の空 token batch を処理し、自動クラスタリングの `max_speakers` 上限を適用します。リアルタイム WebSocket サービスには JSON object を検証する `--hf_overrides` を追加しました。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.19)
前のリリース **FunASR 1.4.18** は中国語とラテン文字が混在する hotword の誤置換、特殊 token 除去後の timestamp 対応、中国語 VAD 分割での文と話者のずれを修正しました。[リリースと検証範囲 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.18)
diff --git a/README_ko.md b/README_ko.md
index a4a8ece06..259840723 100644
--- a/README_ko.md
+++ b/README_ko.md
@@ -111,7 +111,9 @@ checkpoint/revision, 오디오 집합, 하드웨어, 배치, 워밍업, 측정
## 최신 소식
- **MOSS-Transcribe-Diarize**를 FunASR service, Docker, Kubernetes, vLLM/SGLang workflow, FunClip에 통합해 긴 오디오 ASR, timestamp, 익명 speaker label을 한 번에 처리합니다. [MOSS 배포 ->](./docs/moss_transcribe_diarize.md)
-- **FunASR 1.4.21**은 스트리밍 문장부호 처리의 빈 입력이나 공백 입력에서 cache를 유지하고, 캐시된 단어나 embedding 경로를 건드리지 않은 채 빈 결과를 반환합니다. 비어 있지 않은 추론은 변경되지 않습니다. [수정 #3824](https://github.com/modelscope/FunASR/pull/3824). `python -m pip install -U "funasr==1.4.21"`로 설치할 수 있습니다. [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
+- **FunASR 1.4.23**은 `lfr_m=lfr_n=1` 온라인 프런트엔드의 프레임 위치와 오프라인 VAD의 불필요한 빈 마지막 청크를 수정합니다. `token_list`를 지정하면 SentencePiece `ids2tokens`는 모델 토큰을 반환하며, 텍스트 변환에는 `tokens2text`를 사용합니다. 미지정 시 기존 텍스트 디코딩은 유지됩니다. [수정 #3849](https://github.com/modelscope/FunASR/pull/3849), [수정 #3783](https://github.com/modelscope/FunASR/pull/3783). `python -m pip install -U "funasr==1.4.23"`로 설치할 수 있습니다. [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.23)
+ **FunASR 1.4.22**는 호출별 `merge_length_s`를 반영하고 현재 발화 구간 길이에 따라 동적 VAD 무음 임계값을 프레임마다 갱신합니다. 고정 임계값 모드도 유지합니다. [수정 #3841](https://github.com/modelscope/FunASR/pull/3841), [수정 #3833](https://github.com/modelscope/FunASR/pull/3833). [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.22)
+ **FunASR 1.4.21**은 스트리밍 문장부호 처리의 빈 입력에서 cache를 유지합니다. 비어 있지 않은 추론은 변경되지 않습니다. [수정 #3824](https://github.com/modelscope/FunASR/pull/3824). [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
**FunASR 1.4.20**은 `spk_model` 사용 시 선택적 화자 의존성 가져오기가 실패하면 원래 예외를 연결한 `ImportError`를 제공합니다. ASR 전용 구성은 계속 사용할 수 있습니다. [수정 #3819](https://github.com/modelscope/FunASR/pull/3819). [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.20)
**FunASR 1.4.19**는 Paraformer의 빈 token batch를 처리하고 자동 클러스터링에 `max_speakers` 상한을 적용합니다. 실시간 WebSocket 서비스에 JSON object를 검증하는 `--hf_overrides`도 추가했습니다. [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.19)
이전 릴리스 **FunASR 1.4.18**은 중국어와 라틴 문자가 섞인 hotword의 잘못된 치환, 특수 token 제거 후 timestamp 정렬, 중국어 VAD 분할의 문장 및 화자 불일치를 수정했습니다. [릴리스 및 검증 범위 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.18)
diff --git a/README_zh.md b/README_zh.md
index eccb4f714..1c3199b47 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -157,7 +157,9 @@ checkpoint/revision、音频集、硬件、批量大小、预热、计时范围
## 最新动态
- **MOSS-Transcribe-Diarize** 已接入 FunASR 服务、Docker、Kubernetes、vLLM/SGLang 工作流和 FunClip,一次完成长音频转写、时间戳与匿名说话人标注。[部署 MOSS ->](./docs/moss_transcribe_diarize_zh.md)
-- **FunASR 1.4.21** 在流式标点收到空白或空输入时保留 cache,并在不访问缓存词或 embedding 路径的情况下返回空结果;非空推理保持不变。[修复 #3824](https://github.com/modelscope/FunASR/pull/3824)。安装:`python -m pip install -U "funasr==1.4.21"`。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
+- **FunASR 1.4.23** 修复 `lfr_m=lfr_n=1` 在线前端的帧游标和离线 VAD 多余的空尾块。提供 `token_list` 时,SentencePiece 的 `ids2tokens` 返回模型词元,请用 `tokens2text` 转为文本;未提供 `token_list` 时保持旧有文本解码行为。[修复 #3849](https://github.com/modelscope/FunASR/pull/3849)、[修复 #3783](https://github.com/modelscope/FunASR/pull/3783)。安装:`python -m pip install -U "funasr==1.4.23"`。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.23)
+ **FunASR 1.4.22** 支持逐调用覆盖 `merge_length_s`,并根据当前语音段时长逐帧更新动态 VAD 静音阈值,同时保留固定阈值模式。[修复 #3841](https://github.com/modelscope/FunASR/pull/3841)、[修复 #3833](https://github.com/modelscope/FunASR/pull/3833)。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.22)
+ **FunASR 1.4.21** 在流式标点收到空白或空输入时保留 cache;非空推理保持不变。[修复 #3824](https://github.com/modelscope/FunASR/pull/3824)。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)
**FunASR 1.4.20** 在指定 `spk_model` 而可选说话人依赖导入失败时,通过带原始异常链的 `ImportError` 提供诊断,同时保留仅 ASR 的构造路径。[修复 #3819](https://github.com/modelscope/FunASR/pull/3819)。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.20)
**FunASR 1.4.19** 修复 Paraformer 空 token 批次处理,使自动聚类遵守 `max_speakers` 上限,并为实时 WebSocket 服务新增带 JSON 对象校验的 `--hf_overrides` 参数。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.19)
上一版本 **FunASR 1.4.18** 修复跨中英文热词误替换、过滤特殊符号后的时间戳对齐,以及中文 VAD 分段的句子与说话人错位。[发布说明与验证范围 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.18)
diff --git a/docs/python_api.md b/docs/python_api.md
index 160a66273..f95a5ba27 100644
--- a/docs/python_api.md
+++ b/docs/python_api.md
@@ -42,7 +42,7 @@ The wrapper accepts `**kwargs`, not a universally validated parameter schema. A
| `batch_size_s` | `300` | VAD-path batching budget in seconds, based on longest padded segment times segment count, not a hard file-duration limit. CPU VAD processing uses single-segment batches. |
| `batch_size_threshold_s` | `60` | VAD batching threshold for segment duration, not a VAD splitting setting. |
| `merge_vad` | `False` | Enables merging detected VAD regions for this call. |
-| `merge_length_s` | `15` | VAD merge target read before call overrides are merged into ASR config. Set this at construction in this checkout. |
+| `merge_length_s` | `15` | Target length in seconds for merged VAD regions when `merge_vad=True`. A per-call value overrides the construction value. |
| `cache` | No wrapper-managed session | Pass a caller-owned dictionary on every streaming call. It is not a hub/model download cache. |
| `sentence_timestamp` | `False` | Requests sentence records in the VAD pipeline; depends on available timestamps and punctuation, with limited VAD fallback. |
| `return_spk_res` | `True` | Includes clustered speaker results when VAD and a speaker model are configured. `False` does not avoid construction or segment embedding computation. |
diff --git a/docs/python_api_zh.md b/docs/python_api_zh.md
index 42d9a4a13..1f70e6d90 100644
--- a/docs/python_api_zh.md
+++ b/docs/python_api_zh.md
@@ -42,7 +42,7 @@
| `batch_size_s` | `300` | VAD 路径的秒级批处理预算,按最长补齐片段长度乘以片段数量估算,不是文件时长硬上限。CPU 的 VAD 路径按单片段处理。 |
| `batch_size_threshold_s` | `60` | VAD 片段参与组批的时长阈值,不是 VAD 切分参数。 |
| `merge_vad` | `False` | 为本次调用启用 VAD 区域合并。 |
-| `merge_length_s` | `15` | VAD 合并目标,在本次参数合并进 ASR 配置之前读取。当前代码版本应在构建时设置它。 |
+| `merge_length_s` | `15` | `merge_vad=True` 时合并后 VAD 片段的目标长度,单位为秒。本次调用传入的值会覆盖构建时设置的值。 |
| `cache` | 包装层不维护会话 | 每次流式调用都传入调用方持有的字典。它不是模型下载缓存。 |
| `sentence_timestamp` | `False` | 请求 VAD 流程输出句子记录,依赖实际时间戳与标点,部分情形可回退到 VAD 区域。 |
| `return_spk_res` | `True` | 配置 VAD 与说话人模型时输出聚类结果。设为 `False` 并不能省去模型构建或片段说话人特征计算。 |
diff --git a/docs/repository_roles.md b/docs/repository_roles.md
index 911a2b655..96aa91ed4 100644
--- a/docs/repository_roles.md
+++ b/docs/repository_roles.md
@@ -5,7 +5,7 @@ This document explains the responsibility boundaries, user entry points, and iss
> **Directional roadmap, not a release promise.**
> This document records shipped capabilities and active work, but does not commit to
> future version numbers or dates. The current Python release is
-> [`funasr==1.4.21`](https://github.com/modelscope/FunASR/releases/tag/v1.4.21).
+> [`funasr==1.4.23`](https://github.com/modelscope/FunASR/releases/tag/v1.4.23).
> Any future breaking release still requires a maintainer-approved milestone and
> migration plan.
diff --git a/docs/repository_roles_zh.md b/docs/repository_roles_zh.md
index 6a847f170..c0214965e 100644
--- a/docs/repository_roles_zh.md
+++ b/docs/repository_roles_zh.md
@@ -4,7 +4,7 @@
> **方向性路线图,不是版本承诺。**
> 本文档记录已交付能力与正在推进的工作,但不承诺未来版本号或日期。当前 Python
-> 版本是 [`funasr==1.4.21`](https://github.com/modelscope/FunASR/releases/tag/v1.4.21)。
+> 版本是 [`funasr==1.4.23`](https://github.com/modelscope/FunASR/releases/tag/v1.4.23)。
> 任何未来的 breaking release 仍需 maintainer 确认 milestone 与迁移方案。
---
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 8ec9fcbce..b8320dfa7 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -1043,10 +1043,10 @@ class AutoModel:
# FIX(gcf): concat the vad clips for sense vocie model for better aed
if cfg.get("merge_vad", False):
+ # Call options are merged into kwargs only after this point.
+ merge_length_s = cfg.get("merge_length_s", kwargs.get("merge_length_s", 15))
for i in range(len(res)):
- res[i]["value"] = merge_vad(
- res[i]["value"], kwargs.get("merge_length_s", 15) * 1000
- )
+ res[i]["value"] = merge_vad(res[i]["value"], merge_length_s * 1000)
# step.2 compute asr model
model = self.model
diff --git a/funasr/frontends/wav_frontend.py b/funasr/frontends/wav_frontend.py
index 593cc1446..cb671e6df 100644
--- a/funasr/frontends/wav_frontend.py
+++ b/funasr/frontends/wav_frontend.py
@@ -488,6 +488,7 @@ class WavFrontendOnline(nn.Module):
lfr_splice_frame_idxs = []
for i in range(batch_size):
mat = input[i, : input_lengths[i], :]
+ lfr_splice_frame_idx = mat.size(0)
if self.lfr_m != 1 or self.lfr_n != 1:
# update self.lfr_splice_cache in self.apply_lfr
# mat, self.lfr_splice_cache[i], lfr_splice_frame_idx = self.apply_lfr(mat, self.lfr_m, self.lfr_n, self.lfr_splice_cache[i],
diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.py
index d8537ed20..b59862e6b 100644
--- a/funasr/models/fsmn_vad_streaming/model.py
+++ b/funasr/models/fsmn_vad_streaming/model.py
@@ -353,6 +353,7 @@ class Stats(object):
self.output_data_buf_offset = 0
self.frame_probs = []
self.max_end_sil_frame_cnt_thresh = max_end_sil_frame_cnt_thresh
+ self.silence_schedule = None
self.speech_noise_thres = speech_noise_thres
self.scores = None
self.max_time_out = False
@@ -1004,7 +1005,9 @@ class FsmnVADStreaming(nn.Module):
audio_sample = torch.cat((cache["prev_samples"], audio_sample_list[0]))
- n = int(len(audio_sample) // chunk_stride_samples + int(_is_final))
+ n = len(audio_sample) // chunk_stride_samples
+ if _is_final and len(audio_sample) % chunk_stride_samples:
+ n += 1
m = int(len(audio_sample) % chunk_stride_samples * (1 - int(_is_final)))
segments = []
# Keep explicit fixed-threshold requests from being overwritten by the dynamic schedule.
@@ -1012,9 +1015,11 @@ class FsmnVADStreaming(nn.Module):
"dynamic_silence", kwargs.get("max_end_silence_time") is None
)
silence_schedule = kwargs.get("silence_schedule", DEFAULT_SILENCE_SCHEDULE)
- speech_to_sil_ms = self.vad_opts.speech_to_sil_time_thres
- accumulated_ms = cache.get("_dynamic_accumulated_ms", 0)
- in_speech = cache.get("_dynamic_in_speech", False)
+ # The schedule is looked up per frame in DetectOneFrame from the current segment's duration.
+ if "stats" in cache:
+ cache["stats"].silence_schedule = silence_schedule if dynamic_silence else None
+ if dynamic_silence:
+ cache["stats"].speech_noise_thres = 0.5
frontend_inference_kwargs = {}
if getattr(frontend, "supports_aligned_waveforms", False):
frontend_inference_kwargs["return_waveform"] = True
@@ -1023,21 +1028,6 @@ class FsmnVADStreaming(nn.Module):
kwargs["is_final"] = _is_final and i == n - 1
audio_sample_i = audio_sample[i * chunk_stride_samples : (i + 1) * chunk_stride_samples]
- # Apply dynamic silence threshold (only accumulate while in speech)
- if dynamic_silence and "stats" in cache:
- vad_state = cache["stats"].vad_state_machine
- # VadStateMachine: 2=kVadInStateStartPointNotDetected, 3=kVadInStateInSpeechSegment
- if vad_state.value == 2 or in_speech: # kVadInStateInSpeechSegment
- accumulated_ms += chunk_size
- in_speech = True
- for limit_ms, silence_ms in silence_schedule:
- if accumulated_ms <= limit_ms:
- cache["stats"].max_end_sil_frame_cnt_thresh = max(silence_ms - speech_to_sil_ms, 0)
- cache["stats"].speech_noise_thres = 0.5
- break
- cache["_dynamic_accumulated_ms"] = accumulated_ms
- cache["_dynamic_in_speech"] = in_speech
-
# extract fbank feats
speech, speech_lengths = extract_fbank(
[audio_sample_i],
@@ -1074,11 +1064,6 @@ class FsmnVADStreaming(nn.Module):
segments_i = self.forward(**batch)
if len(segments_i) > 0:
segments.extend(*segments_i)
- if dynamic_silence:
- accumulated_ms = 0
- in_speech = False
- cache["_dynamic_accumulated_ms"] = 0
- cache["_dynamic_in_speech"] = False
cache["prev_samples"] = audio_sample[-m:] if m > 0 else torch.empty(0)
if _is_final:
@@ -1264,6 +1249,17 @@ class FsmnVADStreaming(nn.Module):
cur_frm_idx - self.LatencyFrmNumAtStartPoint(cache=cache), cache=cache
)
elif cache["stats"].vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
+ silence_schedule = getattr(cache["stats"], "silence_schedule", None)
+ if silence_schedule is not None:
+ segment_ms = (
+ cur_frm_idx - cache["stats"].confirmed_start_frame + 1
+ ) * frm_shift_in_ms
+ for limit_ms, silence_ms in silence_schedule:
+ if segment_ms <= limit_ms:
+ cache["stats"].max_end_sil_frame_cnt_thresh = max(
+ silence_ms - self.vad_opts.speech_to_sil_time_thres, 0
+ )
+ break
if (
cache["stats"].continous_silence_frame_count * frm_shift_in_ms
>= cache["stats"].max_end_sil_frame_cnt_thresh
diff --git a/funasr/tokenizer/sentencepiece_tokenizer.py b/funasr/tokenizer/sentencepiece_tokenizer.py
index acba22c2a..1c76c8984 100644
--- a/funasr/tokenizer/sentencepiece_tokenizer.py
+++ b/funasr/tokenizer/sentencepiece_tokenizer.py
@@ -13,11 +13,11 @@ from funasr.register import tables
class SentencepiecesTokenizer(BaseTokenizer):
def __init__(self, bpemodel: Union[Path, str], **kwargs):
"""Initialize SentencepiecesTokenizer.
-
- Args:
- bpemodel: TODO.
- **kwargs: Additional keyword arguments.
- """
+
+ Args:
+ bpemodel: TODO.
+ **kwargs: Additional keyword arguments.
+ """
super().__init__(**kwargs)
self.bpemodel = str(bpemodel)
# NOTE(kamo):
@@ -41,39 +41,39 @@ class SentencepiecesTokenizer(BaseTokenizer):
def text2tokens(self, line: str) -> List[str]:
"""Text2tokens.
-
- Args:
- line: TODO.
- """
+
+ Args:
+ line: TODO.
+ """
self._build_sentence_piece_processor()
return self.sp.EncodeAsPieces(line)
def tokens2text(self, tokens: Iterable[str]) -> str:
"""Tokens2text.
-
- Args:
- tokens: TODO.
- """
+
+ Args:
+ tokens: TODO.
+ """
self._build_sentence_piece_processor()
return self.sp.DecodePieces(list(tokens))
def encode(self, line: str, **kwargs) -> List[int]:
"""Encode.
-
- Args:
- line: TODO.
- **kwargs: Additional keyword arguments.
- """
+
+ Args:
+ line: TODO.
+ **kwargs: Additional keyword arguments.
+ """
self._build_sentence_piece_processor()
return self.sp.EncodeAsIds(line)
def decode(self, line: List[int], **kwargs):
"""Decode.
-
- Args:
- line: TODO.
- **kwargs: Additional keyword arguments.
- """
+
+ Args:
+ line: TODO.
+ **kwargs: Additional keyword arguments.
+ """
self._build_sentence_piece_processor()
return self.sp.DecodeIds(line)
@@ -81,20 +81,17 @@ class SentencepiecesTokenizer(BaseTokenizer):
"""Get vocab size."""
return self.sp.GetPieceSize()
- def ids2tokens(self, *args, **kwargs):
- """Ids2tokens.
-
- Args:
- *args: Variable positional arguments.
- **kwargs: Additional keyword arguments.
- """
- return self.decode(*args, **kwargs)
+ def ids2tokens(self, line, **kwargs):
+ """Return model-vocabulary pieces, or legacy decoded text without a token list."""
+ if getattr(self, "token_list", None) is not None:
+ return super().ids2tokens(line)
+ return self.decode(line, **kwargs)
def tokens2ids(self, *args, **kwargs):
"""Tokens2ids.
-
- Args:
- *args: Variable positional arguments.
- **kwargs: Additional keyword arguments.
- """
+
+ Args:
+ *args: Variable positional arguments.
+ **kwargs: Additional keyword arguments.
+ """
return self.encode(*args, **kwargs)
diff --git a/funasr/version.txt b/funasr/version.txt
index 8cc8fe024..580ef4b6c 100644
--- a/funasr/version.txt
+++ b/funasr/version.txt
@@ -1 +1 @@
-1.4.21
+1.4.23
diff --git a/runtime/python/libtorch/README.md b/runtime/python/libtorch/README.md
index 9136bd74f..00808b6ba 100644
--- a/runtime/python/libtorch/README.md
+++ b/runtime/python/libtorch/README.md
@@ -21,12 +21,23 @@ python -m funasr.export.export_model --model-name damo/speech_paraformer-large_a
## Install the `funasr_torch`
+`funasr_torch` is a separate distribution from `funasr`; upgrading the root
+package alone does not update these TorchScript wrappers. Version 0.1.5
+includes the Contextual/Seaco Paraformer hotword batching and empty
+hotword-file handling fix (#3838). This is a runtime wrapper fix, not a
+model-weight update or an ASR accuracy claim.
+
+Version 0.1.6 makes the standalone CMVN reader accept blank lines and
+raise a file-specific error for missing or truncated statistics (#3842).
+Valid float64 statistics and shipped feature output are unchanged. This is
+a CPU parser/package fix, not a model-weight or ASR accuracy update.
+
install from pip
```shell
-pip install -U "funasr_torch>=0.1.4"
+pip install -U "funasr_torch>=0.1.6"
# For the users in China, you could install with the command:
-# pip install -U funasr_torch -i https://mirror.sjtu.edu.cn/pypi/web/simple
+# pip install -U "funasr_torch>=0.1.6" -i https://mirror.sjtu.edu.cn/pypi/web/simple
```
or install from source code
diff --git a/runtime/python/libtorch/funasr_torch/utils/frontend.py b/runtime/python/libtorch/funasr_torch/utils/frontend.py
index 743ee5e5f..2772ad95d 100644
--- a/runtime/python/libtorch/funasr_torch/utils/frontend.py
+++ b/runtime/python/libtorch/funasr_torch/utils/frontend.py
@@ -125,25 +125,29 @@ class WavFrontend:
self,
) -> np.ndarray:
with open(self.cmvn_file, "r", encoding="utf-8") as f:
- lines = f.readlines()
+ lines = [line for line in f.readlines() if line.split()]
means_list = []
vars_list = []
for i in range(len(lines)):
line_item = lines[i].split()
if line_item[0] == "<AddShift>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
add_shift_line = line_item[3 : (len(line_item) - 1)]
means_list = list(add_shift_line)
continue
elif line_item[0] == "<Rescale>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
rescale_line = line_item[3 : (len(line_item) - 1)]
vars_list = list(rescale_line)
continue
+ if not means_list or not vars_list:
+ raise ValueError(
+ f"No <AddShift>/<Rescale> statistics found in cmvn file: {self.cmvn_file}"
+ )
means = np.array(means_list).astype(np.float64)
vars = np.array(vars_list).astype(np.float64)
cmvn = np.array([means, vars])
diff --git a/runtime/python/libtorch/setup.py b/runtime/python/libtorch/setup.py
index 3ba4610e4..be4b7e3c5 100644
--- a/runtime/python/libtorch/setup.py
+++ b/runtime/python/libtorch/setup.py
@@ -15,7 +15,7 @@ def get_readme():
setuptools.setup(
name="funasr_torch",
- version="0.1.4",
+ version="0.1.6",
platforms="Any",
url="https://github.com/modelscope/FunASR.git",
author="Speech Lab of DAMO Academy, Alibaba Group",
diff --git a/runtime/python/onnxruntime/README.md b/runtime/python/onnxruntime/README.md
index 0b19d0cc5..8794ac4c7 100644
--- a/runtime/python/onnxruntime/README.md
+++ b/runtime/python/onnxruntime/README.md
@@ -7,6 +7,17 @@ does not update the installed ONNX wrappers. Check their versions separately
with `python -m pip show funasr-onnx funasr`; use the source installation below
when a required wrapper fix is newer than the published PyPI package.
+Version 0.4.9 keeps one result per input when offline VAD receives empty or sub-frame
+audio, without invoking ONNX Runtime for zero-frame features (#3846). It also returns
+an empty punctuation chunk while it preserves the streaming cache and avoids
+tokenization and inference (#3847). These are CPU runtime-wrapper fixes, not a
+model-weight or ASR accuracy update.
+
+Version 0.4.8 makes the standalone CMVN reader accept blank lines and
+raise a file-specific error for missing or truncated statistics (#3842).
+Valid float64 statistics and shipped feature output are unchanged. This is
+a CPU parser/package fix, not a model-weight or ASR accuracy update.
+
Version 0.4.7 includes the Contextual/Seaco Paraformer hotword batching fix
(#3838), including empty hotword-file handling. This is a runtime wrapper
fix, not a model-weight update or an ASR accuracy claim.
@@ -28,9 +39,9 @@ and postprocessing fixes, not a model-weight update or an ASR accuracy claim.
install from pip
```shell
-pip install -U "funasr-onnx>=0.4.7"
+pip install -U "funasr-onnx>=0.4.9"
# For the users in China, you could install with the command:
-# pip install -U "funasr-onnx>=0.4.7" -i https://mirror.sjtu.edu.cn/pypi/web/simple
+# pip install -U "funasr-onnx>=0.4.9" -i https://mirror.sjtu.edu.cn/pypi/web/simple
# If you want to export .onnx file, you should install modelscope and funasr
pip install -U modelscope funasr
# For the users in China, you could install with the command:
diff --git a/runtime/python/onnxruntime/funasr_onnx/punc_bin.py b/runtime/python/onnxruntime/funasr_onnx/punc_bin.py
index 7fa62d44d..bfcb6f80b 100644
--- a/runtime/python/onnxruntime/funasr_onnx/punc_bin.py
+++ b/runtime/python/onnxruntime/funasr_onnx/punc_bin.py
@@ -99,6 +99,8 @@ class CT_Transformer:
split_text = self.code_mix_split_words_jieba(text)
else:
split_text = code_mix_split_words(text)
+ if not split_text:
+ return "", []
split_text_id = self.converter.tokens2ids(split_text)
mini_sentences = split_to_mini_sentence(split_text, split_size)
mini_sentences_id = split_to_mini_sentence(split_text_id, split_size)
@@ -196,6 +198,9 @@ class CT_Transformer_VadRealtime(CT_Transformer):
cache_key = "cache"
assert cache_key in param_dict
cache = param_dict[cache_key]
+ # An empty chunk adds no text and must not reprocess pending words.
+ if not text.strip():
+ return "", [], cache if cache is not None else []
if cache is not None and len(cache) > 0:
precache = "".join(cache)
else:
diff --git a/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py b/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
index f150ccbf8..4218718b6 100644
--- a/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
+++ b/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
@@ -142,24 +142,26 @@ def load_cmvn(cmvn_file: Union[str, Path]) -> np.ndarray:
raise FileNotFoundError("cmvn file not exits")
with open(cmvn_file, "r", encoding="utf-8") as f:
- lines = f.readlines()
+ lines = [line for line in f.readlines() if line.split()]
means_list = []
vars_list = []
for i in range(len(lines)):
line_item = lines[i].split()
if line_item[0] == "<AddShift>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
add_shift_line = line_item[3 : (len(line_item) - 1)]
means_list = list(add_shift_line)
continue
elif line_item[0] == "<Rescale>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
rescale_line = line_item[3 : (len(line_item) - 1)]
vars_list = list(rescale_line)
continue
+ if not means_list or not vars_list:
+ raise ValueError(f"No <AddShift>/<Rescale> statistics found in cmvn file: {cmvn_file}")
means = np.array(means_list).astype(np.float64)
vars = np.array(vars_list).astype(np.float64)
cmvn = np.array([means, vars])
diff --git a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
index cb9ba3d0d..5d9d8fd4e 100644
--- a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
+++ b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
@@ -103,6 +103,8 @@ class Fsmn_vad:
end_idx = beg_idx + 1
waveform = waveform_list[beg_idx:end_idx]
feats, feats_len = self.extract_feat(waveform)
+ if feats_len[0] == 0:
+ continue
waveform = np.array(waveform)
param_dict = kwargs.get("param_dict", dict())
in_cache = param_dict.get("in_cache", list())
@@ -183,8 +185,13 @@ class Fsmn_vad:
def extract_feat(self, waveform_list: List[np.ndarray]) -> Tuple[np.ndarray, np.ndarray]:
feats, feats_len = [], []
for waveform in waveform_list:
- speech, _ = self.frontend.fbank(waveform)
- feat, feat_len = self.frontend.lfr_cmvn(speech)
+ speech, speech_len = self.frontend.fbank(waveform)
+ if speech_len == 0:
+ # LFR needs a first frame; keep the empty input at its original position.
+ feat = np.empty((0, speech.shape[1] * self.frontend.lfr_m), dtype=speech.dtype)
+ feat_len = 0
+ else:
+ feat, feat_len = self.frontend.lfr_cmvn(speech)
feats.append(feat)
feats_len.append(feat_len)
diff --git a/runtime/python/onnxruntime/setup.py b/runtime/python/onnxruntime/setup.py
index b61aa96c6..b5236ffc1 100644
--- a/runtime/python/onnxruntime/setup.py
+++ b/runtime/python/onnxruntime/setup.py
@@ -13,7 +13,7 @@ def get_readme():
MODULE_NAME = "funasr_onnx"
-VERSION_NUM = "0.4.7"
+VERSION_NUM = "0.4.9"
setuptools.setup(
name=MODULE_NAME,
diff --git a/tests/test_fsmn_vad_dynamic_silence.py b/tests/test_fsmn_vad_dynamic_silence.py
index f396614f9..d8f39c345 100644
--- a/tests/test_fsmn_vad_dynamic_silence.py
+++ b/tests/test_fsmn_vad_dynamic_silence.py
@@ -55,13 +55,83 @@ class TestFsmnVadDynamicSilence(unittest.TestCase):
cache = self._run_inference(max_end_silence_time=300)
self.assertEqual(cache["stats"].max_end_sil_frame_cnt_thresh, 200)
- self.assertNotIn("_dynamic_accumulated_ms", cache)
+ self.assertIsNone(cache["stats"].silence_schedule)
+ self.assertEqual(cache["stats"].speech_noise_thres, 0.6)
def test_explicit_dynamic_silence_still_enables_schedule(self):
cache = self._run_inference(max_end_silence_time=300, dynamic_silence=True)
- self.assertEqual(cache["stats"].max_end_sil_frame_cnt_thresh, 1900)
- self.assertEqual(cache["_dynamic_accumulated_ms"], 1000)
+ self.assertEqual(cache["stats"].silence_schedule, vad_model.DEFAULT_SILENCE_SCHEDULE)
+ self.assertEqual(cache["stats"].speech_noise_thres, 0.5)
+ # The threshold is no longer advanced per chunk; DetectOneFrame applies the schedule.
+ self.assertEqual(cache["stats"].max_end_sil_frame_cnt_thresh, 200)
+ self.assertNotIn("_dynamic_accumulated_ms", cache)
+
+
+class TestFsmnVadDynamicSilencePerFrame(unittest.TestCase):
+ """Drive DetectOneFrame with scripted window-detector transitions."""
+
+ def _end_frames(self, speech_ms, silence_ms, schedule, start_frame=0, fixed_thresh_ms=650):
+ vad = vad_model.FsmnVADStreaming.__new__(vad_model.FsmnVADStreaming)
+ torch.nn.Module.__init__(vad)
+ vad.vad_opts = vad_model.VADXOptions()
+ ends = []
+
+ def on_voice_end(end_frame, fake_result, is_last_frame, cache=None):
+ ends.append(end_frame)
+
+ def reset_detection(cache=None):
+ cache["stats"].confirmed_start_frame = -1
+ cache["stats"].vad_state_machine = vad_model.VadStateMachine.kVadInStateStartPointNotDetected
+
+ vad.OnVoiceEnd = on_voice_end
+ vad.PopDataToOutputBuf = lambda *args, **kwargs: None
+ vad.PopDataBufTillFrame = lambda *args, **kwargs: None
+ vad.LatencyFrmNumAtStartPoint = lambda cache=None: 0
+ vad.ResetDetection = reset_detection
+
+ stats = vad_model.Stats(
+ sil_pdf_ids=[0], max_end_sil_frame_cnt_thresh=fixed_thresh_ms, speech_noise_thres=0.5
+ )
+ stats.silence_schedule = schedule
+ speech_frames, silence_frames = speech_ms // 10, silence_ms // 10
+ changes = (
+ [vad_model.AudioChangeState.kChangeStateSil2Speech]
+ + [vad_model.AudioChangeState.kChangeStateSpeech2Speech] * (speech_frames - 1)
+ + [vad_model.AudioChangeState.kChangeStateSpeech2Sil]
+ + [vad_model.AudioChangeState.kChangeStateSil2Sil] * silence_frames
+ )
+ script = iter(changes)
+ cache = {
+ "stats": stats,
+ "windows_detector": SimpleNamespace(DetectOneFrame=lambda *a, **k: next(script)),
+ }
+ for i in range(len(changes)):
+ vad.DetectOneFrame(vad_model.FrameState.kFrameStateSpeech, start_frame + i, False, cache=cache)
+ return [e - start_frame for e in ends]
+
+ def test_schedule_follows_current_segment_duration(self):
+ schedule = vad_model.DEFAULT_SILENCE_SCHEDULE
+ # Thresholds below are the schedule values minus speech_to_sil_time_thres (150 ms).
+ # 5 s into a segment: 2000 ms -> 1850 ms, so 900 ms of silence does not end it.
+ self.assertEqual(self._end_frames(5000, 900, schedule), [])
+ # 15 s into a segment: 1000 ms -> 850 ms, so 900 ms ends it and 700 ms does not.
+ self.assertEqual(len(self._end_frames(15000, 900, schedule)), 1)
+ self.assertEqual(self._end_frames(15000, 700, schedule), [])
+ # 55 s into a segment: 200 ms -> 50 ms, so 300 ms ends it.
+ self.assertEqual(len(self._end_frames(55000, 300, schedule)), 1)
+
+ def test_schedule_does_not_depend_on_absolute_position(self):
+ schedule = vad_model.DEFAULT_SILENCE_SCHEDULE
+ for speech_ms, silence_ms in [(5000, 900), (15000, 900), (55000, 300)]:
+ at_zero = self._end_frames(speech_ms, silence_ms, schedule)
+ shifted = self._end_frames(speech_ms, silence_ms, schedule, start_frame=6000)
+ self.assertEqual(at_zero, shifted)
+
+ def test_without_schedule_threshold_stays_fixed(self):
+ # 650 ms fixed threshold: 700 ms of silence ends a 5 s segment, 600 ms does not.
+ self.assertEqual(len(self._end_frames(5000, 700, None)), 1)
+ self.assertEqual(self._end_frames(5000, 600, None), [])
if __name__ == "__main__":
diff --git a/tests/test_funasr_onnx_installed.py b/tests/test_funasr_onnx_installed.py
index babcf53a5..4b3c254ee 100644
--- a/tests/test_funasr_onnx_installed.py
+++ b/tests/test_funasr_onnx_installed.py
@@ -38,7 +38,7 @@ ENTRYPOINTS = (
class InstalledOnnxPackageTest(unittest.TestCase):
def test_installed_distribution_version_and_origin(self):
distribution = importlib.metadata.distribution('funasr-onnx')
- self.assertEqual(distribution.version, '0.4.7')
+ self.assertEqual(distribution.version, '0.4.9')
self.assertEqual(
Path(funasr_onnx.__file__).resolve().parent,
Path(distribution.locate_file('funasr_onnx')).resolve(),
diff --git a/tests/test_funasr_onnx_punctuation_empty_input.py b/tests/test_funasr_onnx_punctuation_empty_input.py
new file mode 100644
index 000000000..901a228bc
--- /dev/null
+++ b/tests/test_funasr_onnx_punctuation_empty_input.py
@@ -0,0 +1,116 @@
+"""Run with PYTHONPATH=runtime/python/onnxruntime python -m unittest discover
+-s tests -p test_funasr_onnx_punctuation_empty_input.py (standalone ONNX dependencies).
+"""
+
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import Mock
+
+import numpy as np
+
+from funasr_onnx.punc_bin import CT_Transformer, CT_Transformer_VadRealtime
+from funasr_onnx.utils.utils import TokenIDConverter, code_mix_split_words_jieba
+
+
+EMPTY_TEXTS = ("", " ", "\t\n", " \t\r\n\u3000")
+PUNC_LIST = ["<unk>", "_", ",", "。", "?", "、"]
+
+
+def make_model(model_class, segmenter=None):
+ # Exercise real tokenization, chunking and cache handling without model downloads.
+ model = model_class.__new__(model_class)
+ model.seg_jieba = segmenter is not None
+ model.code_mix_split_words_jieba = segmenter
+ model.converter = TokenIDConverter(["<unk>", "你", "好", "世", "界", "hello", "world"])
+ model.punc_list = PUNC_LIST
+ model.period = PUNC_LIST.index("。")
+
+ def infer(tokens, lengths, *masks):
+ logits = np.zeros((*tokens.shape, len(PUNC_LIST)), dtype=np.float32)
+ logits[:, :, PUNC_LIST.index("_")] = 1
+ return [logits]
+
+ model.infer = Mock(side_effect=infer)
+ return model
+
+
+class OnnxPunctuationEmptyInputTest(unittest.TestCase):
+ @classmethod
+ def setUpClass(cls):
+ with tempfile.TemporaryDirectory() as directory:
+ dictionary = Path(directory) / "jieba_usr_dict"
+ dictionary.write_text("你好 100\n世界 100\n", encoding="utf-8")
+ cls.segmenter = staticmethod(code_mix_split_words_jieba(str(dictionary)))
+
+ def test_offline_empty_skips_inference(self):
+ for segmenter in (None, self.segmenter):
+ model = make_model(CT_Transformer, segmenter)
+ for text in EMPTY_TEXTS:
+ with self.subTest(jieba=segmenter is not None, text=text):
+ self.assertEqual(model(text), ("", []))
+ model.infer.assert_not_called()
+
+ def test_offline_nonempty_around_empty(self):
+ for segmenter in (None, self.segmenter):
+ for text, expected in (("你好", "你好。"), ("hello world", "hello world。")):
+ with self.subTest(jieba=segmenter is not None, text=text):
+ model = make_model(CT_Transformer, segmenter)
+ before = model(text)
+ self.assertEqual(before[0], expected)
+ self.assertEqual(before[1][-1], model.period)
+ for empty in EMPTY_TEXTS:
+ self.assertEqual(model(empty), ("", []))
+ self.assertEqual(model(text), before)
+ self.assertEqual(model.infer.call_count, 2)
+
+ def test_streaming_empty_preserves_cache(self):
+ for initial_cache in (None, [], ["你", "好"], ["hello", " world"]):
+ with self.subTest(cache=initial_cache):
+ model = make_model(CT_Transformer_VadRealtime)
+ params = {"cache": initial_cache}
+ expected = None if initial_cache is None else initial_cache.copy()
+ for text in EMPTY_TEXTS * 2:
+ result = model(text, params)
+ self.assertEqual(result, ("", [], expected or []))
+ self.assertIs(params["cache"], initial_cache)
+ self.assertEqual(params["cache"], expected)
+ if initial_cache is not None:
+ self.assertIs(result[2], initial_cache)
+ model.infer.assert_not_called()
+
+ def test_streaming_continuity_across_empty_chunks(self):
+ for chunks in (("你好", "世界"), ("hello", "world"), ("你好 hello", "world 世界")):
+ for split_size in (2, 20):
+ with self.subTest(chunks=chunks, split_size=split_size):
+ model = make_model(CT_Transformer_VadRealtime)
+ reference = make_model(CT_Transformer_VadRealtime)
+ params, reference_params = {"cache": []}, {"cache": []}
+ for chunk in chunks:
+ for empty in EMPTY_TEXTS:
+ cache = params["cache"]
+ calls = model.infer.call_count
+ self.assertEqual(model(empty, params), ("", [], cache))
+ self.assertIs(params["cache"], cache)
+ self.assertEqual(model.infer.call_count, calls)
+ self.assertEqual(
+ model(chunk, params, split_size=split_size),
+ reference(chunk, reference_params, split_size=split_size),
+ )
+ self.assertEqual(params, reference_params)
+ self.assertEqual(model.infer.call_count, reference.infer.call_count)
+ for actual, expected in zip(
+ model.infer.call_args_list, reference.infer.call_args_list
+ ):
+ for actual_arg, expected_arg in zip(actual.args, expected.args):
+ np.testing.assert_array_equal(actual_arg, expected_arg)
+
+ def test_streaming_requires_cache_key_even_for_empty_text(self):
+ model = make_model(CT_Transformer_VadRealtime)
+ with self.assertRaises(AssertionError):
+ model("", {})
+ model.infer.assert_not_called()
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/tests/test_funasr_onnx_release.py b/tests/test_funasr_onnx_release.py
index 34a9fdea4..160c24aa4 100644
--- a/tests/test_funasr_onnx_release.py
+++ b/tests/test_funasr_onnx_release.py
@@ -7,7 +7,7 @@ import unittest
ROOT = Path(__file__).resolve().parents[1]
PACKAGE_ROOT = ROOT / "runtime" / "python" / "onnxruntime"
SETUP_PATH = PACKAGE_ROOT / "setup.py"
-EXPECTED_VERSION = "0.4.7"
+EXPECTED_VERSION = "0.4.9"
REQUIREMENT_NAME_PATTERN = re.compile(r"^\s*([A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?)")
@@ -44,7 +44,7 @@ def normalized_requirement_name(requirement):
class FunASROnnxReleaseContractTest(unittest.TestCase):
- def test_release_version_is_0_4_7(self):
+ def test_release_version_is_0_4_9(self):
self.assertEqual(assigned_literal(read_setup_tree(), "VERSION_NUM"), EXPECTED_VERSION)
def test_standalone_license_preserves_repository_notice(self):
@@ -58,6 +58,12 @@ class FunASROnnxReleaseContractTest(unittest.TestCase):
self.assertIn(f'assert version("funasr-onnx") == "{EXPECTED_VERSION}"', workflow)
self.assertIn(f"name: funasr-onnx-{EXPECTED_VERSION}", workflow)
+ def test_workflow_collects_empty_punctuation_regression(self):
+ workflow = (ROOT / ".github/workflows/test-funasr-onnx-package.yml").read_text()
+ test_path = "tests/test_funasr_onnx_punctuation_empty_input.py"
+ self.assertEqual(workflow.count(test_path), 3)
+ self.assertIn("python " + test_path, workflow)
+
def test_readme_names_the_standalone_fix_release(self):
readme = (PACKAGE_ROOT / "README.md").read_text()
self.assertIn(f'pip install -U "funasr-onnx>={EXPECTED_VERSION}"', readme)
@@ -69,6 +75,12 @@ class FunASROnnxReleaseContractTest(unittest.TestCase):
self.assertIn("#3831", readme)
self.assertIn("#3838", readme)
self.assertIn("hotword", readme)
+ self.assertIn("#3842", readme)
+ self.assertIn("CMVN", readme)
+ self.assertIn("#3846", readme)
+ self.assertIn("#3847", readme)
+ self.assertIn("empty or sub-frame", readme)
+ self.assertIn("preserves the streaming cache", readme)
def test_runtime_dependencies_keep_onnx_install_torch_free(self):
requirements = setup_keyword_literal(read_setup_tree(), "install_requires")
diff --git a/tests/test_funasr_onnx_vad_multiple_inputs.py b/tests/test_funasr_onnx_vad_multiple_inputs.py
index fd2478114..a5cbaff48 100644
--- a/tests/test_funasr_onnx_vad_multiple_inputs.py
+++ b/tests/test_funasr_onnx_vad_multiple_inputs.py
@@ -4,6 +4,7 @@ import importlib.util
import os
import tempfile
import unittest
+from unittest.mock import Mock
import numpy as np
@@ -44,10 +45,10 @@ def fake_infer(inputs):
return scores, inputs[1:]
-def make_model(batch_size=1):
+def make_model(batch_size=1, **frontend_conf):
# Keep __call__, WavFrontend and E2EVadModel real; no model files.
model = Fsmn_vad.__new__(Fsmn_vad)
- model.frontend = WavFrontend(cmvn_file=None, **FRONTEND_CONF)
+ model.frontend = WavFrontend(cmvn_file=None, **{**FRONTEND_CONF, **frontend_conf})
model.batch_size = batch_size
# E2EVadModel's defaults equal the template's model_conf.
model.vad_scorer_config = {}
@@ -99,6 +100,80 @@ class FsmnVadMultipleInputsTest(unittest.TestCase):
result = model([self.wav[name] for name in names])
self.assertEqual(result, [SEGMENTS[name] for name in names])
+ def test_short_inputs_and_frame_boundary(self):
+ for fs, frame_length, lfr_m in (
+ (16000, 25, 5),
+ (8000, 25, 5),
+ (16000, 20, 5),
+ (16000, 25, 1),
+ ):
+ window = int(fs * frame_length / 1000)
+ for batch_size in (1, 2, 3):
+ model = make_model(batch_size, fs=fs, frame_length=frame_length, lfr_m=lfr_m)
+ model.infer = Mock(side_effect=fake_infer)
+ for samples in (0, 1, window // 2, window - 1, window, window + 1, 2 * window):
+ audio = np.zeros(samples, dtype=np.float32)
+ path = os.path.join(os.path.dirname(self.wav["a"]), "short.wav")
+ soundfile.write(path, audio, fs)
+ for audio_in in (audio, path):
+ with self.subTest(
+ fs=fs,
+ frame_length=frame_length,
+ lfr_m=lfr_m,
+ batch_size=batch_size,
+ samples=samples,
+ input_type=type(audio_in).__name__,
+ ):
+ model.infer.reset_mock()
+ self.assertEqual(model(audio_in), [[]])
+ self.assertEqual(model.infer.call_count, int(samples >= window))
+
+ def test_short_files_keep_their_input_positions(self):
+ for samples in (0, 160, 320, 399):
+ path = os.path.join(os.path.dirname(self.wav["a"]), "short.wav")
+ soundfile.write(path, np.zeros(samples, dtype=np.float32), 16000)
+ paths = {**self.wav, "short": path}
+ expected = {**SEGMENTS, "short": []}
+ for batch_size in (1, 2, 3):
+ model = make_model(batch_size)
+ model.infer = Mock(side_effect=fake_infer)
+ for names in (
+ ["short", "a"],
+ ["a", "short"],
+ ["a", "short", "b"],
+ ["short", "short"],
+ ["short", "b", "short"],
+ ):
+ with self.subTest(samples=samples, batch_size=batch_size, names=names):
+ model.infer.reset_mock()
+ result = model([paths[name] for name in names])
+ self.assertEqual(result, [expected[name] for name in names])
+ self.assertEqual(
+ model.infer.call_count, sum(name != "short" for name in names)
+ )
+ self.assertEqual(len({id(segments) for segments in result}), len(result))
+
+ def test_extract_feat_with_empty_frames(self):
+ for lfr_m, lfr_n in ((1, 1), (5, 1), (3, 2)):
+ model = make_model(lfr_m=lfr_m, lfr_n=lfr_n)
+ empty = np.zeros(0, dtype=np.float32)
+ short = np.zeros(399, dtype=np.float32)
+ normal = np.zeros(1600, dtype=np.float32)
+ expected, expected_len = model.extract_feat([normal])
+ for waveforms, lengths in (
+ ([empty], [0]),
+ ([short], [0]),
+ ([empty, short, normal], [0, 0, expected_len[0]]),
+ ):
+ with self.subTest(lfr_m=lfr_m, lfr_n=lfr_n, lengths=lengths):
+ feats, feats_len = model.extract_feat(waveforms)
+ np.testing.assert_array_equal(feats_len, lengths)
+ self.assertEqual(feats.shape, (len(waveforms), max(lengths), 80 * lfr_m))
+ self.assertEqual(feats.dtype, np.float32)
+ if len(waveforms) > 1:
+ np.testing.assert_array_equal(feats[-1], expected[0])
+ self.assertFalse(feats[:-1].any())
+
if __name__ == "__main__":
unittest.main()
diff --git a/tests/test_funasr_torch_package.py b/tests/test_funasr_torch_package.py
index ce39c8613..04cdcf01f 100644
--- a/tests/test_funasr_torch_package.py
+++ b/tests/test_funasr_torch_package.py
@@ -119,12 +119,12 @@ def test_torch_package_does_not_require_unrelated_runtime_engines(distributions)
}
assert not names & {"torch-quant", "onnxruntime", "onnxruntime-gpu"}
-def test_distribution_version_excludes_empty_published_release(distributions):
+def test_distribution_version_includes_cmvn_parser_fix(distributions):
from packaging.version import Version
wheel, sdist = distributions
version = metadata(wheel)["Version"]
- assert Version(version) >= Version("0.1.4")
+ assert Version(version) >= Version("0.1.6")
with tarfile.open(sdist, "r:gz") as archive:
member = next(item for item in archive.getmembers()
if item.name.count("/") == 1 and item.name.endswith("/PKG-INFO"))
diff --git a/tests/test_merge_vad_length.py b/tests/test_merge_vad_length.py
new file mode 100644
index 000000000..f14fc952d
--- /dev/null
+++ b/tests/test_merge_vad_length.py
@@ -0,0 +1,81 @@
+"""Regression tests for the merge_length_s option of AutoModel.generate()."""
+
+import unittest
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+
+
+VAD_SEGMENTS = [[0, 4000], [5000, 9000], [10000, 14000]]
+
+
+class TestMergeVadLength(unittest.TestCase):
+ def _make_auto_model(self, **init_kwargs):
+ from funasr.auto.auto_model import AutoModel
+
+ am = AutoModel.__new__(AutoModel)
+ am.model = MagicMock()
+ am.vad_model = MagicMock()
+ am.punc_model = None
+ am.punc_kwargs = {}
+ am.spk_model = None
+ am.spk_kwargs = {}
+ am.cb_model = None
+ am.spk_mode = None
+ am.vad_kwargs = {}
+ am.kwargs = {
+ "batch_size_s": 300,
+ "batch_size_threshold_s": 60,
+ "device": "cpu",
+ "disable_pbar": True,
+ "frontend": MagicMock(fs=16000),
+ "fs": 16000,
+ **init_kwargs,
+ }
+ am._reset_runtime_configs = MagicMock()
+
+ def mock_inference(data, input_len=None, model=None, kwargs=None, **cfg):
+ if model is am.vad_model:
+ return [{"key": "utt", "value": [list(s) for s in VAD_SEGMENTS]}]
+ return [{"text": "x"} for _ in data]
+
+ am.inference = MagicMock(side_effect=mock_inference)
+ return am
+
+ @patch("funasr.auto.auto_model.slice_padding_audio_samples")
+ @patch("funasr.auto.auto_model.load_audio_text_image_video")
+ @patch("funasr.auto.auto_model.prepare_data_iterator")
+ def _asr_segments(self, am, mock_prep, mock_load, mock_slice, **cfg):
+ mock_prep.return_value = (["utt"], [np.zeros(16000 * 14, dtype=np.float32)])
+ mock_load.return_value = np.zeros(16000 * 14, dtype=np.float32)
+ mock_slice.side_effect = lambda speech, speech_lengths, segments: (
+ [np.zeros(16000, dtype=np.float32) for _ in segments],
+ [16000 for _ in segments],
+ )
+
+ am.inference_with_vad("dummy_input", merge_vad=True, **cfg)
+
+ segments = []
+ for call in mock_slice.call_args_list:
+ segments.extend(seg for seg, _ in call.args[2])
+ return sorted(segments)
+
+ def test_default_merge_length_is_15_seconds(self):
+ am = self._make_auto_model()
+ self.assertEqual(self._asr_segments(am), [[0, 14000]])
+
+ def test_generate_merge_length_s_is_applied(self):
+ am = self._make_auto_model()
+ self.assertEqual(self._asr_segments(am, merge_length_s=10), [[0, 9000], [9000, 14000]])
+
+ def test_construction_merge_length_s_is_applied(self):
+ am = self._make_auto_model(merge_length_s=10)
+ self.assertEqual(self._asr_segments(am), [[0, 9000], [9000, 14000]])
+
+ def test_generate_merge_length_s_overrides_construction_value(self):
+ am = self._make_auto_model(merge_length_s=10)
+ self.assertEqual(self._asr_segments(am, merge_length_s=15), [[0, 14000]])
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/tests/test_release_version_contract.py b/tests/test_release_version_contract.py
index 57623d690..621d3ed83 100644
--- a/tests/test_release_version_contract.py
+++ b/tests/test_release_version_contract.py
@@ -4,16 +4,18 @@ import pytest
ROOT = Path(__file__).resolve().parents[1]
-RELEASE_VERSION = "1.4.21"
+RELEASE_VERSION = "1.4.23"
README_FILES = ("README.md", "README_zh.md", "README_ja.md", "README_ko.md")
REPOSITORY_ROLE_FILES = ("docs/repository_roles.md", "docs/repository_roles_zh.md")
@pytest.mark.parametrize("name", README_FILES)
-def test_release_summary_preserves_streaming_punctuation_cache(name):
+def test_release_summary_covers_no_lfr_and_tokenizer_vocabulary(name):
summaries = [line for line in (ROOT / name).read_text(encoding="utf-8").splitlines()
if f"**FunASR {RELEASE_VERSION}**" in line]
assert len(summaries) == 1
- assert "cache" in summaries[0].lower()
- assert "https://github.com/modelscope/FunASR/pull/3824" in summaries[0]
+ assert "lfr_m=lfr_n=1" in summaries[0]
+ assert "token_list" in summaries[0]
+ assert "https://github.com/modelscope/FunASR/pull/3849" in summaries[0]
+ assert "https://github.com/modelscope/FunASR/pull/3783" in summaries[0]
diff --git a/tests/test_runtime_cmvn.py b/tests/test_runtime_cmvn.py
new file mode 100644
index 000000000..97889e2fc
--- /dev/null
+++ b/tests/test_runtime_cmvn.py
@@ -0,0 +1,105 @@
+"""Exercise the standalone runtime frontends without model weights or engines."""
+
+import importlib.util
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+CMVN = "<AddShift> 2 2\n<LearnRateCoef> 1 [ 1 -2 ]\n<Rescale> 2 2\n<LearnRateCoef> 1 [ 0.5 2 ]\n"
+
+
[email protected](params=["onnxruntime/funasr_onnx", "libtorch/funasr_torch"])
+def frontend_class(request):
+ # Loading the frontend directly keeps these tests independent of the
+ # optional inference engines and exercises each package's real constructor.
+ path = ROOT / "runtime/python" / request.param / "utils/frontend.py"
+ spec = importlib.util.spec_from_file_location("runtime_cmvn_frontend", path)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module.WavFrontend
+
+
+def test_valid_cmvn_preserves_statistics_and_normalized_features(frontend_class, tmp_path):
+ path = tmp_path / "valid.mvn"
+ path.write_text(CMVN, encoding="utf-8")
+ frontend = frontend_class(cmvn_file=str(path))
+
+ assert frontend.cmvn.dtype == np.float64
+ np.testing.assert_array_equal(frontend.cmvn, [[1, -2], [0.5, 2]])
+ features = np.array([[0, 2], [-2, 4]], dtype=np.float32)
+ np.testing.assert_array_equal(frontend.apply_cmvn(features), [[0.5, 0], [-0.5, 4]])
+
+
[email protected]("placement", ["leading", "trailing", "between_sections", "after_tags"])
+def test_cmvn_accepts_blank_lines(frontend_class, tmp_path, placement):
+ blank = " \t\n\n"
+ if placement == "leading":
+ content = blank + CMVN
+ elif placement == "trailing":
+ content = CMVN + blank
+ elif placement == "between_sections":
+ content = CMVN.replace("<Rescale>", blank + "<Rescale>")
+ else:
+ content = CMVN.replace("<LearnRateCoef>", blank + "<LearnRateCoef>")
+ path = tmp_path / "blank lines.mvn"
+ path.write_text(content, encoding="utf-8")
+
+ frontend = frontend_class(cmvn_file=str(path))
+
+ np.testing.assert_array_equal(frontend.cmvn, [[1, -2], [0.5, 2]])
+
+
+ "content",
+ [
+ "",
+ " \t\n",
+ "<Nnet>\n</Nnet>\n",
+ "<AddShift> 2 2\n",
+ "<Rescale> 2 2\n",
+ "<AddShift> 2 2\n<LearnRateCoef> 1 [ 1 -2 ]\n",
+ "<Rescale> 2 2\n<LearnRateCoef> 1 [ 0.5 2 ]\n",
+ CMVN.replace("\n", " "),
+ ],
+ ids=[
+ "empty",
+ "blank",
+ "no_stats",
+ "truncated_shift",
+ "truncated_scale",
+ "shift_only",
+ "scale_only",
+ "compact",
+ ],
+)
+def test_incomplete_cmvn_fails_during_construction(frontend_class, tmp_path, content):
+ path = tmp_path / "incomplete.mvn"
+ path.write_text(content, encoding="utf-8")
+
+ with pytest.raises(ValueError) as error:
+ frontend_class(cmvn_file=str(path))
+
+ assert str(path) in str(error.value)
+
+
+def test_shipped_cmvn_is_unchanged_by_blank_lines(frontend_class, tmp_path):
+ shipped = ROOT / "runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/am.mvn"
+ original = frontend_class(cmvn_file=str(shipped), lfr_m=7, lfr_n=6, dither=0.0)
+ path = tmp_path / "shipped-with-blanks.mvn"
+ path.write_text("\n" + "\n \t\n".join(shipped.read_text().splitlines()) + "\n\n")
+ edited = frontend_class(cmvn_file=str(path), lfr_m=7, lfr_n=6, dither=0.0)
+
+ assert original.cmvn.shape == (2, 560)
+ assert np.isfinite(original.cmvn).all()
+ np.testing.assert_array_equal(edited.cmvn, original.cmvn)
+ # Real filterbank + LFR + CMVN on a deterministic waveform, not a model stub.
+ waveform = np.sin(2 * np.pi * 440 * np.arange(1600) / 16000).astype(np.float32) * 0.1
+ original_features, _ = original.fbank(waveform)
+ edited_features, _ = edited.fbank(waveform)
+ original_features, original_length = original.lfr_cmvn(original_features)
+ edited_features, edited_length = edited.lfr_cmvn(edited_features)
+ assert edited_length == original_length
+ np.testing.assert_array_equal(edited_features, original_features)
diff --git a/tests/test_sentencepiece_tokenizer.py b/tests/test_sentencepiece_tokenizer.py
new file mode 100644
index 000000000..f3a4349ea
--- /dev/null
+++ b/tests/test_sentencepiece_tokenizer.py
@@ -0,0 +1,162 @@
+"""Model vocabulary IDs must reach callers as pieces, not BPE-decoded text."""
+
+import json
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+import sentencepiece as spm
+import torch
+
+from funasr.models.paraformer.model import Paraformer
+from funasr.tokenizer.build_tokenizer import build_tokenizer
+from funasr.tokenizer.sentencepiece_tokenizer import SentencepiecesTokenizer
+
+
[email protected](scope="module")
+def bpe_model(tmp_path_factory):
+ prefix = tmp_path_factory.mktemp("sentencepiece") / "tiny"
+ spm.SentencePieceTrainer.train(
+ sentence_iterator=iter(
+ ["hello world", "hello there", "he tried to think how it could be", "你好 world"] * 20
+ ),
+ model_prefix=str(prefix),
+ model_type="bpe",
+ vocab_size=48,
+ user_defined_symbols=["<|en|>"],
+ shuffle_input_sentence=False,
+ num_threads=1,
+ minloglevel=2,
+ )
+ return str(prefix.with_suffix(".model"))
+
+
+def vocabulary(bpe_model, reordered=True):
+ sp = spm.SentencePieceProcessor(model_file=bpe_model)
+ pieces = [sp.id_to_piece(i) for i in range(sp.get_piece_size())]
+ return list(reversed(pieces)) if reordered else pieces
+
+
[email protected]("storage", ["list", "tuple", "json", "txt"])
+def test_reordered_vocabulary_decodes_expected_text(bpe_model, tmp_path, storage):
+ tokens = vocabulary(bpe_model)
+ token_list = tokens
+ if storage == "tuple":
+ token_list = tuple(tokens)
+ elif storage in ("json", "txt"):
+ path = tmp_path / ("tokens." + storage)
+ content = json.dumps(tokens) if storage == "json" else "\n".join(tokens) + "\n"
+ path.write_text(content, encoding="utf-8")
+ token_list = str(path)
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=token_list)
+ pieces = tokenizer.text2tokens("hello world")
+ model_ids = [tokens.index(piece) for piece in pieces]
+
+ # Establish that the two vocabularies really assign different meanings to IDs.
+ assert tokenizer.decode(model_ids) != "hello world"
+ result = tokenizer.ids2tokens(model_ids)
+
+ assert result == pieces
+ assert tokenizer.tokens2text(result) == "hello world"
+
+
[email protected]("text", ["hello world", "你好 world", "<|en|> hello world", ""])
+def test_matching_vocabulary_preserves_text_and_piece_boundaries(bpe_model, text):
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=vocabulary(bpe_model, False))
+ ids = tokenizer.encode(text)
+
+ pieces = tokenizer.ids2tokens(ids)
+
+ assert pieces == tokenizer.text2tokens(text)
+ assert tokenizer.tokens2text(pieces) == tokenizer.decode(ids) == text
+
+
[email protected]("container", [list, tuple, iter, np.asarray])
+def test_model_ids_accept_base_tokenizer_containers(bpe_model, container):
+ tokens = vocabulary(bpe_model)
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=tokens)
+ ids = [tokens.index(piece) for piece in tokenizer.text2tokens("hello world")]
+
+ assert tokenizer.tokens2text(tokenizer.ids2tokens(container(ids))) == "hello world"
+
+
+def test_special_tokens_and_model_only_ids_are_preserved(bpe_model):
+ tokens = vocabulary(bpe_model) + ["<blank>"]
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=tokens)
+ pieces = ["<s>", "<unk>", "<|en|>", "▁", "▁", "</s>", "<blank>"]
+
+ assert tokenizer.ids2tokens([tokens.index(piece) for piece in pieces]) == pieces
+ assert tokenizer.ids2tokens([]) == []
+
+
+def test_model_ids_require_one_dimensional_numpy_array(bpe_model):
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=vocabulary(bpe_model))
+
+ with pytest.raises(ValueError, match="Must be 1 dim ndarray"):
+ tokenizer.ids2tokens(np.array([[1, 2]]))
+
+
[email protected]("construction", ["omitted", "none", "build_tokenizer"])
+def test_without_vocabulary_preserves_legacy_text_decode(bpe_model, construction):
+ if construction == "build_tokenizer":
+ tokenizer = build_tokenizer("bpe", bpemodel=bpe_model)
+ elif construction == "none":
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=None)
+ else:
+ tokenizer = SentencepiecesTokenizer(bpe_model)
+ ids = tokenizer.encode("hello world")
+
+ assert tokenizer.ids2tokens(ids) == "hello world"
+ assert tokenizer.ids2tokens(line=ids) == "hello world"
+ assert tokenizer.ids2tokens([ids, ids]) == ["hello world", "hello world"]
+ assert tokenizer.ids2tokens([]) == ""
+
+
+def test_encode_decode_and_legacy_keyword_remain_sentencepiece_operations(bpe_model):
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=vocabulary(bpe_model))
+ ids = tokenizer.encode("hello world")
+ model_ids = [tokenizer.token_list.index(p) for p in tokenizer.text2tokens("hello world")]
+
+ assert tokenizer.tokens2ids("hello world") == ids
+ assert tokenizer.decode(ids) == "hello world"
+ assert tokenizer.tokens2text(tokenizer.ids2tokens(line=model_ids, unused=True)) == "hello world"
+
+
+def test_paraformer_inference_uses_model_pieces_and_filters_control_ids(bpe_model, tmp_path):
+ tokens = ["<blank>", "<s>", "</s>"] + [
+ p for p in vocabulary(bpe_model) if p not in ("<s>", "</s>")
+ ]
+ tokenizer = SentencepiecesTokenizer(bpe_model, token_list=tokens)
+ pieces = tokenizer.text2tokens("hello world")
+ ids = [0, 1] + [tokens.index(p) for p in pieces] + [2, 0]
+ scores = torch.full((1, len(ids), len(tokens)), -10.0)
+ scores[0, torch.arange(len(ids)), ids] = 10.0
+ lengths = torch.tensor([len(ids)])
+ speech = torch.zeros(1, len(ids), 2)
+ # Only the acoustic stages are controlled; the actual inference method,
+ # tokenizer, special-ID filtering, text conversion and writer all run.
+ state = SimpleNamespace(
+ beam_search=None,
+ sos=1,
+ eos=2,
+ blank_id=0,
+ encode=lambda *args: (speech, lengths),
+ calc_predictor=lambda *args: (speech, lengths.float(), None, None),
+ cal_decoder_with_predictor=lambda *args: (scores, lengths),
+ )
+
+ results, _ = Paraformer.inference(
+ state,
+ speech,
+ data_lengths=lengths,
+ key=["sample"],
+ tokenizer=tokenizer,
+ data_type="fbank",
+ device="cpu",
+ output_dir=str(tmp_path),
+ )
+
+ assert results == [{"key": "sample", "text": "hello world"}]
+ assert (tmp_path / "1best_recog" / "token").read_text().strip() == (
+ "sample " + " ".join(pieces)
+ )
diff --git a/tests/test_utterance_mvn.py b/tests/test_utterance_mvn.py
index f74201cd0..851ee3886 100644
--- a/tests/test_utterance_mvn.py
+++ b/tests/test_utterance_mvn.py
@@ -194,8 +194,22 @@ def test_saved_config_and_checkpoint_round_trip(tmp_path):
torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+def ieee_fp32_matmul():
+ # Batch shapes can select different TF32 kernels; test the FP32 contract.
+ matmul = torch.backends.cuda.matmul.allow_tf32
+ cudnn = torch.backends.cudnn.allow_tf32
+ try:
+ torch.backends.cuda.matmul.allow_tf32 = False
+ torch.backends.cudnn.allow_tf32 = False
+ yield
+ finally:
+ torch.backends.cuda.matmul.allow_tf32 = matmul
+ torch.backends.cudnn.allow_tf32 = cudnn
+
+
@pytest.mark.parametrize("norm_vars", [False, True])
-def test_real_audio_transformer_inference(device, norm_vars):
+def test_real_audio_transformer_inference(device, norm_vars, ieee_fp32_matmul):
"""Real loader/frontend/encoder/CTC; small random model, no downloads or training."""
audio, fs = sf.read(
Path(__file__).resolve().parents[1] / "runtime/funasr_api/asr_example.wav", dtype="float32"
diff --git a/tests/test_wav_frontend_no_lfr.py b/tests/test_wav_frontend_no_lfr.py
new file mode 100644
index 000000000..a9af79662
--- /dev/null
+++ b/tests/test_wav_frontend_no_lfr.py
@@ -0,0 +1,133 @@
+import pytest
+import torch
+
+from funasr.frontends.wav_frontend import WavFrontendOnline
+from funasr.models.fsmn_vad_streaming import model as vad_model
+from funasr.utils import fbank as kaldi
+
+
+def make_frontend(tmp_path, cmvn):
+ cmvn_file = None
+ if cmvn:
+ cmvn_file = tmp_path / "no_lfr.mvn"
+ cmvn_file.write_text(
+ "<AddShift> 80 80\n<LearnRateCoef> 1 [ "
+ + " ".join(["0.25"] * 80)
+ + " ]\n<Rescale> 80 80\n<LearnRateCoef> 1 [ "
+ + " ".join(["2"] * 80)
+ + " ]\n"
+ )
+ return WavFrontendOnline(cmvn_file=cmvn_file, dither=0.0)
+
+
[email protected]("cmvn", (False, True))
[email protected]("is_final", (False, True))
[email protected]("lengths", ((5, 3, 0), (0, 0)))
+def test_no_lfr_reports_valid_consumed_frames(tmp_path, cmvn, is_final, lengths):
+ frontend = make_frontend(tmp_path, cmvn)
+ inputs = torch.arange(len(lengths) * 7 * 80, dtype=torch.float32).reshape(
+ len(lengths), 7, 80
+ )
+ original = inputs.clone()
+ cache = {}
+
+ feats, actual_lengths, consumed = frontend.forward_lfr_cmvn(
+ inputs, torch.tensor(lengths), is_final=is_final, cache=cache
+ )
+
+ assert actual_lengths.tolist() == list(lengths)
+ assert consumed.tolist() == list(lengths)
+ assert feats.shape == (len(lengths), max(lengths), 80)
+ assert feats.dtype == torch.float32
+ assert cache == {}
+ for index, length in enumerate(lengths):
+ expected = original[index, :length]
+ if cmvn:
+ expected = (expected + 0.25) * 2
+ torch.testing.assert_close(feats[index, :length], expected, rtol=0, atol=0)
+ assert torch.count_nonzero(feats[index, length:]) == 0
+
+
[email protected]("cmvn", (False, True))
[email protected]("empty_final", (False, True))
[email protected]("chunks", ((1600,), (400,) * 4, (160,) * 10, (80, 320, 40, 1160)))
+def test_no_lfr_stream_matches_full_filterbank(tmp_path, cmvn, empty_final, chunks):
+ frontend = make_frontend(tmp_path, cmvn)
+ audio = 0.2 * torch.sin(torch.arange(1600, dtype=torch.float32) * (2 * torch.pi * 440 / 16000))
+ expected = kaldi.fbank(
+ audio.unsqueeze(0) * (1 << 15), num_mel_bins=80, frame_length=25, frame_shift=10,
+ sample_frequency=16000, window_type="hamming", dither=0.0, energy_floor=0.0,
+ )
+ if cmvn:
+ expected = (expected + 0.25) * 2
+ cache, outputs, offset, emitted = {}, [], 0, 0
+ calls = chunks + ((0,) if empty_final else ())
+ for index, size in enumerate(calls):
+ chunk = audio[offset : offset + size]
+ offset += size
+ feats, lengths = frontend(
+ chunk.unsqueeze(0), torch.tensor([size]), cache=cache,
+ is_final=index == len(calls) - 1, return_waveform=True,
+ )
+ count = int(lengths[0]) if lengths.numel() else 0
+ if count:
+ outputs.append(feats[0, :count])
+ start = emitted * 160
+ stop = start + (count - 1) * 160 + 400
+ torch.testing.assert_close(cache["aligned_waveforms"][0], audio[start:stop], rtol=0, atol=0)
+ emitted += count
+ else:
+ assert cache["aligned_waveforms"].numel() == 0
+ assert cache["reserve_waveforms"].numel() == 0
+ assert emitted == expected.shape[0] == 8
+ torch.testing.assert_close(torch.cat(outputs), expected, rtol=1e-5, atol=1e-5)
+
+
[email protected]("size", (0, 160, 399))
[email protected]("empty_final", (False, True))
+def test_no_lfr_short_stream_has_no_frames(tmp_path, size, empty_final):
+ frontend = make_frontend(tmp_path, False)
+ cache = {}
+ for index, length in enumerate((size, 0) if empty_final else (size,)):
+ feats, lengths = frontend(
+ torch.zeros(1, length), torch.tensor([length]), cache=cache,
+ is_final=not empty_final or index == 1, return_waveform=True,
+ )
+ assert feats.numel() == 0
+ assert not lengths.numel() or lengths.sum() == 0
+ assert cache["aligned_waveforms"].numel() == 0
+
+
[email protected]("lfr_m", (1, 5))
[email protected]("offset_ms", (0, 1000))
[email protected]("chunk_ms", (20, 100, 5000))
+def test_offline_vad_finalizes_last_nonempty_chunk(lfr_m, offset_ms, chunk_ms):
+ class ScriptedEncoder(torch.nn.Module):
+ def __init__(self):
+ super().__init__()
+ self.frames = 0
+
+ def forward(self, feats, cache=None):
+ time_ms = (torch.arange(feats.shape[1]) + self.frames) * 10
+ speech = ((time_ms >= 500 + offset_ms) & (time_ms < 1500 + offset_ms)) | (
+ (time_ms >= 2500 + offset_ms) & (time_ms < 3200 + offset_ms)
+ )
+ self.frames += feats.shape[1]
+ probability = torch.where(speech, 0.999, 0.001)
+ return torch.stack((1 - probability, probability), dim=-1).unsqueeze(0)
+
+ model = vad_model.FsmnVADStreaming.__new__(vad_model.FsmnVADStreaming)
+ torch.nn.Module.__init__(model)
+ model.encoder = ScriptedEncoder()
+ model.vad_opts = vad_model.VADXOptions()
+ audio = 0.2 * torch.sin(
+ torch.arange((4000 + offset_ms) * 16, dtype=torch.float32)
+ * (2 * torch.pi * 440 / 16000)
+ )
+ frontend = WavFrontendOnline(lfr_m=lfr_m, lfr_n=1, dither=0.0)
+ result, _ = model.inference(
+ [audio], key=["fixture"], cache={}, frontend=frontend, device="cpu",
+ chunk_size=chunk_ms, is_streaming_input=False, is_final=True, dynamic_silence=True,
+ )
+ assert model.encoder.frames == (4000 + offset_ms) // 10 - 2
+ assert result == [{"key": "fixture", "value": [[240 + offset_ms, 3980 + offset_ms]]}]