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(subtitle): preserve more VAD context by default
zhifu gao483ca541eaac
diff --git a/examples/subtitle/README.md b/examples/subtitle/README.mdindex f173b52e6..7624bbf18 100644--- a/examples/subtitle/README.md+++ b/examples/subtitle/README.md@@ -22,6 +22,9 @@ python generate_subtitle.py audio.wav --device cpu# Preserve raw model sentence boundariespython generate_subtitle.py audio.wav --segment-mode sentence++# Use shorter VAD segments when memory is limited+python generate_subtitle.py audio.wav --max-single-segment-time 30000```## Output Example (SRT)@@ -44,6 +47,7 @@ python generate_subtitle.py audio.wav --segment-mode sentence| `--segment-mode` | readable | Cue grouping: readable or raw sentence boundaries || `--model` | SenseVoiceSmall | ASR model || `--device` | cuda | Device: cuda or cpu |+| `--max-single-segment-time` | 60000 | Maximum VAD segment length in milliseconds; reduce it when memory is limited || `--spk` | off | Add speaker labels || `--lang` | auto | Language hint || `-o` | input.srt | Output path |@@ -52,6 +56,10 @@ Readable mode joins only adjacent short or continuation cues within boundedgap, duration, length, and speaker limits. It does not rewrite recognized textor punctuation.+The default 60-second VAD limit gives subtitle generation more context at+speech boundaries. For memory-constrained machines, lower it explicitly (for+example, `--max-single-segment-time 30000`).+## Install```bashdiff --git a/examples/subtitle/generate_subtitle.py b/examples/subtitle/generate_subtitle.pyindex bfa21c6b6..1caeddd3d 100644--- a/examples/subtitle/generate_subtitle.py+++ b/examples/subtitle/generate_subtitle.py@@ -73,6 +73,13 @@ def main():)parser.add_argument("--model", default="iic/SenseVoiceSmall")parser.add_argument("--device", default="cuda")+ parser.add_argument(+ "--max-single-segment-time",+ type=int,+ default=60000,+ metavar="MS",+ help="Maximum VAD segment length in milliseconds (default: 60000)",+ )parser.add_argument("--spk", action="store_true", help="Include speaker labels")parser.add_argument("--lang", default="auto")args = parser.parse_args()@@ -88,7 +95,7 @@ def main():from funasr import AutoModelkwargs = {"model": args.model, "vad_model": "fsmn-vad", "punc_model": "ct-punc",- "vad_kwargs": {"max_single_segment_time": 30000},+ "vad_kwargs": {"max_single_segment_time": args.max_single_segment_time},"device": args.device, "disable_update": True}if args.spk:kwargs["spk_model"] = "cam++"diff --git a/tests/test_generate_subtitle.py b/tests/test_generate_subtitle.pyindex a7779e935..c4d4a3d1a 100644--- a/tests/test_generate_subtitle.py+++ b/tests/test_generate_subtitle.py@@ -68,6 +68,7 @@ def test_generate_subtitle_requests_sentence_timestamps_and_writes_segmented_srtmodule.main()assert captured["model_kwargs"]["punc_model"] == "ct-punc"+ assert captured["model_kwargs"]["vad_kwargs"] == {"max_single_segment_time": 60000}assert captured["generate_kwargs"]["language"] == "zh"assert captured["generate_kwargs"]["sentence_timestamp"] is Trueassert captured["generate_kwargs"]["output_timestamp"] is True