Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
fix(auto): honor merge_length_s passed to generate()
Mohammad Hijjawi099c2f4404be
diff --git a/docs/python_api.md b/docs/python_api.mdindex 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.mdindex 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/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex 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 aedif 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 modelmodel = self.modeldiff --git a/tests/test_merge_vad_length.py b/tests/test_merge_vad_length.pynew file mode 100644index 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()