Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

bug fix

维石af1c0d0e30fc
diff --git a/runtime/python/libtorch/funasr_torch/sensevoice_bin.py b/runtime/python/libtorch/funasr_torch/sensevoice_bin.py
index 8de1c54c8..4f793406d 100644
--- a/runtime/python/libtorch/funasr_torch/sensevoice_bin.py
+++ b/runtime/python/libtorch/funasr_torch/sensevoice_bin.py
@@ -170,15 +170,19 @@ class SenseVoiceSmall:
torch.tensor(_language_list).to(self.device),
torch.tensor(_textnorm_list).to(self.device),
)
- # support batch_size=1 only currently
- x = ctc_logits[0, : encoder_out_lens[0].item(), :]
- yseq = x.argmax(dim=-1)
- yseq = torch.unique_consecutive(yseq, dim=-1)
-
- mask = yseq != self.blank_id
- token_int = yseq[mask].tolist()
-
- asr_res.append(self.tokenizer.decode(token_int))
+ for b in range(feats.shape[0]):
+ # back to torch.Tensor
+ if isinstance(ctc_logits, np.ndarray):
+ ctc_logits = torch.from_numpy(ctc_logits).float()
+ # support batch_size=1 only currently
+ x = ctc_logits[b, : encoder_out_lens[b].item(), :]
+ yseq = x.argmax(dim=-1)
+ yseq = torch.unique_consecutive(yseq, dim=-1)
+
+ mask = yseq != self.blank_id
+ token_int = yseq[mask].tolist()
+
+ asr_res.append(self.tokenizer.decode(token_int))
return asr_res