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: solve problems in sensevoice_bin.py related to argmax and unique, as mentioned in issue #2331 (#2332)
majic31ae7aff2e9cf6
diff --git a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.pyindex 9cfe46f81..6a06ed105 100644--- a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py@@ -183,8 +183,10 @@ class SenseVoiceSmall:# ctc_logits = torch.from_numpy(ctc_logits).float()# support batch_size=1 only currentlyx = ctc_logits[b, : encoder_out_lens[b].item(), :]- yseq = x.argmax(dim=-1)- yseq = np.unique(yseq)+ yseq = np.argmax(x, axis=-1)+ # Use np.diff and np.where instead of torch.unique_consecutive.+ mask = np.concatenate(([True], np.diff(yseq) != 0))+ yseq = yseq[mask]mask = yseq != self.blank_idtoken_int = yseq[mask].tolist()