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: add super().__init__() to SinusoidalPositionEncoder classes; add VAD streaming test
游雁bff79be1a2d1
diff --git a/funasr/models/sense_voice/model.py b/funasr/models/sense_voice/model.pyindex 2dbb7e150..30a9e42b9 100644--- a/funasr/models/sense_voice/model.py+++ b/funasr/models/sense_voice/model.py@@ -26,7 +26,7 @@ class SinusoidalPositionEncoder(torch.nn.Module):""" """def __init__(self, d_model=80, dropout_rate=0.1):- pass+ super().__init__()def encode(self, positions: torch.Tensor = None, depth: int = None, dtype: torch.dtype = torch.float32diff --git a/funasr/models/transformer/embedding.py b/funasr/models/transformer/embedding.pyindex 4797df05d..9ef2f4b0e 100644--- a/funasr/models/transformer/embedding.py+++ b/funasr/models/transformer/embedding.py@@ -384,7 +384,7 @@ class SinusoidalPositionEncoder(torch.nn.Module):""" """def __init__(self, d_model=80, dropout_rate=0.1):- pass+ super().__init__()def encode(self, positions: torch.Tensor = None, depth: int = None, dtype: torch.dtype = torch.float32@@ -417,7 +417,7 @@ class StreamSinusoidalPositionEncoder(torch.nn.Module):""" """def __init__(self, d_model=80, dropout_rate=0.1):- pass+ super().__init__()def encode(self, positions: torch.Tensor = None, depth: int = None, dtype: torch.dtype = torch.float32diff --git a/tests_models/test_fsmn_vad_streaming.py b/tests_models/test_fsmn_vad_streaming.pynew file mode 100644index 000000000..07ba46ca3--- /dev/null+++ b/tests_models/test_fsmn_vad_streaming.py@@ -0,0 +1,98 @@+#!/usr/bin/env python3+"""Test FSMN-VAD Streaming: chunk-by-chunk voice activity detection"""+import sys+import time+import os++def main():+ import soundfile+ from funasr import AutoModel++ print("[FSMN-VAD-Streaming] Loading model...")+ t0 = time.time()+ model = AutoModel(+ model="iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",+ device="cpu",+ disable_update=True,+ disable_pbar=True,+ )+ print("[FSMN-VAD-Streaming] Model loaded in %.1fs" % (time.time() - t0))++ wav_file = os.path.join(model.model_path, "example/vad_example.wav")+ speech, sample_rate = soundfile.read(wav_file)+ chunk_size = 200 # ms+ chunk_stride = int(chunk_size * sample_rate / 1000)++ total_chunk_num = int((len(speech) - 1) / chunk_stride + 1)+ print("[FSMN-VAD-Streaming] Audio: %.2fs, %d chunks of %dms" % (+ len(speech) / sample_rate, total_chunk_num, chunk_size))++ print("[FSMN-VAD-Streaming] Running streaming inference...")+ t0 = time.time()++ cache = {}+ all_events = []+ for i in range(total_chunk_num):+ speech_chunk = speech[i * chunk_stride:(i + 1) * chunk_stride]+ is_final = i == total_chunk_num - 1+ res = model.generate(+ input=speech_chunk,+ cache=cache,+ is_final=is_final,+ chunk_size=chunk_size,+ )+ if res[0]["value"]:+ all_events.extend(res[0]["value"])++ print("[FSMN-VAD-Streaming] Inference done in %.1fs" % (time.time() - t0))++ # Parse streaming VAD events into complete segments+ # Streaming output: [beg, -1] = speech start, [-1, end] = speech end, [beg, end] = complete+ complete_segments = []+ pending_start = None+ for event in all_events:+ if event[0] >= 0 and event[1] == -1:+ pending_start = event[0]+ elif event[0] == -1 and event[1] >= 0:+ if pending_start is not None:+ complete_segments.append([pending_start, event[1]])+ pending_start = None+ elif event[0] >= 0 and event[1] >= 0:+ complete_segments.append(event)++ print("[FSMN-VAD-Streaming] Raw events: %d, Complete segments: %d" % (+ len(all_events), len(complete_segments)))+ print("[FSMN-VAD-Streaming] Segments: %s" % complete_segments)++ if not complete_segments:+ print("[FSMN-VAD-Streaming] FAILED - no complete segments")+ return 1++ # Verify segments have valid ranges+ for seg in complete_segments:+ if seg[1] <= seg[0]:+ print("[FSMN-VAD-Streaming] FAILED - invalid segment: %s" % seg)+ return 1++ # Verify consistency: run again with fresh cache+ cache2 = {}+ all_events2 = []+ for i in range(total_chunk_num):+ speech_chunk = speech[i * chunk_stride:(i + 1) * chunk_stride]+ is_final = i == total_chunk_num - 1+ res = model.generate(+ input=speech_chunk, cache=cache2, is_final=is_final, chunk_size=chunk_size,+ )+ if res[0]["value"]:+ all_events2.extend(res[0]["value"])++ if all_events != all_events2:+ print("[FSMN-VAD-Streaming] FAILED - inconsistent across sessions")+ return 1++ print("[FSMN-VAD-Streaming] Consistency: 2 sessions identical")+ print("[FSMN-VAD-Streaming] PASSED")+ return 0++if __name__ == "__main__":+ sys.exit(main())