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: remove repetition_penalty from vLLM services (fixes #2950 CUDA assert)
游雁b8d736ef8496
diff --git a/examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py b/examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.pyindex b52a1dd1d..5c58bc5ab 100644--- a/examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py+++ b/examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py@@ -318,7 +318,6 @@ class RealtimeASRSession:hotwords=self.asr_kwargs.get("hotwords"),language=self.asr_kwargs.get("language"),max_new_tokens=200,- repetition_penalty=1.3,)text = results[0]["text"] if results else ""text = _clean_asr_text(text)@@ -359,7 +358,6 @@ class RealtimeASRSession:hotwords=self.asr_kwargs.get("hotwords"),language=self.asr_kwargs.get("language"),max_new_tokens=512,- repetition_penalty=1.3,)text = results[0]["text"] if results else ""text = _clean_asr_text(text)diff --git a/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py b/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.pyindex 1820f9517..6a0ab1a95 100644--- a/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py+++ b/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py@@ -32,6 +32,20 @@ warnings.filterwarnings('ignore')logging.basicConfig(level=logging.INFO, format='%(asctime)s [%(levelname)s] %(message)s')logger = logging.getLogger(__name__)+def truncate_repetition(text, min_repeat_len=3, max_repeats=3):+ """Detect and truncate repetitive patterns in ASR output."""+ if not text or len(text) < 20:+ return text+ n = len(text)+ for length in range(min_repeat_len, min(n // max_repeats, 30)):+ for start in range(n - length * max_repeats):+ chunk = text[start:start + length]+ if text[start:start + length * max_repeats] == chunk * max_repeats:+ return text[:start + length]+ return text+++try:from fastapi import FastAPI, File, UploadFile, Form, WebSocket, WebSocketDisconnectfrom fastapi.responses import JSONResponse@@ -107,7 +121,7 @@ def process_audio(audio_data, sr=16000, language=None, hotwords=None,return {"text": "", "segments": [], "duration": len(audio_data) / sr}# vLLM batch ASR- gen_kwargs = {"max_new_tokens": 500, "repetition_penalty": 1.3}+ gen_kwargs = {"max_new_tokens": 500}if language:gen_kwargs["language"] = languageif hotwords:@@ -120,6 +134,7 @@ def process_audio(audio_data, sr=16000, language=None, hotwords=None,full_text_parts = []for i, (r, (start_ms, end_ms)) in enumerate(zip(results, seg_times)):+ r["text"] = truncate_repetition(r["text"])seg_info = {"text": r["text"],"start": start_ms / 1000,