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(server): report detected transcription language (#3400)
zhifu gao8d9b34bebc71
diff --git a/funasr/bin/_server_app.py b/funasr/bin/_server_app.pyindex 1dc6fdced..64e821da3 100644--- a/funasr/bin/_server_app.py+++ b/funasr/bin/_server_app.py@@ -26,6 +26,27 @@ except ImportError:logger = logging.getLogger("funasr.server")+_LANGUAGE_TAG_RE = re.compile(r"<\|(zh|en|yue|ja|ko)\|>")+++def extract_language_from_asr_text(text):+ """Extract a SenseVoice language code before special tokens are removed."""+ if not isinstance(text, str):+ return None+ match = _LANGUAGE_TAG_RE.search(text)+ return match.group(1) if match else None+++def resolve_transcription_language(requested_language, result):+ """Prefer the caller's language hint, then backend detection, else unknown."""+ if requested_language and requested_language.strip().lower() != "auto":+ return requested_language+ detected_language = result.get("language")+ if isinstance(detected_language, str) and detected_language:+ return detected_language+ return "unknown"++def _split_text_for_openai_segments(text: str, max_chars: int = 80):"""Split unsegmented ASR text into readable OpenAI-compatible cues."""text = text.strip()@@ -255,7 +276,9 @@ def create_app(device: str = "cuda", preload_model: str = "auto", model_path: stif language:kwargs["language"] = languageresult = model.generate(**kwargs)- text = re.sub(r'<\|[^|]*\|>', '', result[0]["text"]).strip()+ raw_text = result[0]["text"]+ detected_language = extract_language_from_asr_text(raw_text)+ text = re.sub(r'<\|[^|]*\|>', '', raw_text).strip()segments = []if "sentence_info" in result[0]:for s in result[0]["sentence_info"]:@@ -267,7 +290,12 @@ def create_app(device: str = "cuda", preload_model: str = "auto", model_path: st})if not segments and text:segments = build_openai_fallback_segments(text, duration)- return {"text": text, "segments": segments, "duration": duration}+ return {+ "text": text,+ "segments": segments,+ "duration": duration,+ "language": detected_language,+ }# Pre-loadif app.state.model_path:@@ -322,7 +350,7 @@ def create_app(device: str = "cuda", preload_model: str = "auto", model_path: stif response_format == "verbose_json":return JSONResponse({"task": "transcribe",- "language": language or "zh",+ "language": resolve_transcription_language(language, result),"duration": result.get("duration", 0),"text": result["text"],"segments": [diff --git a/tests/test_server_app_openai_segments.py b/tests/test_server_app_openai_segments.pyindex f29fd14ae..fadbdaa2a 100644--- a/tests/test_server_app_openai_segments.py+++ b/tests/test_server_app_openai_segments.py@@ -103,6 +103,29 @@ def test_fallback_segments_keep_short_text_single_cue(monkeypatch):]+def test_extract_language_from_sensevoice_text(monkeypatch):+ module = load_server_app(monkeypatch)++ assert module.extract_language_from_asr_text("<|en|><|NEUTRAL|><|Speech|>hello") == "en"+ assert module.extract_language_from_asr_text("<|yue|> nei hou") == "yue"+ assert module.extract_language_from_asr_text("plain transcript") is None+++def test_resolve_transcription_language_prefers_request_then_detection(monkeypatch):+ module = load_server_app(monkeypatch)++ assert module.resolve_transcription_language("ja", {"language": "en"}) == "ja"+ assert module.resolve_transcription_language("auto", {"language": "en"}) == "en"+ assert module.resolve_transcription_language(None, {"language": "ko"}) == "ko"+ assert module.resolve_transcription_language(None, {}) == "unknown"+++def test_resolve_transcription_language_does_not_default_to_chinese(monkeypatch):+ module = load_server_app(monkeypatch)++ assert module.resolve_transcription_language(None, {}) != "zh"++def test_default_fun_asr_nano_uses_requested_modelscope_hub(monkeypatch):module = load_server_app(monkeypatch)DummyAutoModel = install_dummy_funasr(monkeypatch)