speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

simple streaming

Browse Source
This commit is contained in:
志浩 2024-09-13 16:24:06 +08:00
parent 441b997f19
commit 4a8cb6f0c4
1 changed files with 37 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
funasr/models/llm_asr/model.py
View File

@ -3130,6 +3130,8 @@ class LLMASRXvecSlotTTS(nn.Module):
_text = f"<|endofprompt|><|sil|>{text}" + ("<|sil|>" if is_last else "") _text = f"<|endofprompt|><|sil|>{text}" + ("<|sil|>" if is_last else "")
text_token = self.tts_tokenizer_warpper(_text) text_token = self.tts_tokenizer_warpper(_text)
cur_token, feat, wav = None, None, None
if len(text_token) > tts_text_chunk_size:
text_token = torch.tensor([text_token], dtype=torch.long, device=device) text_token = torch.tensor([text_token], dtype=torch.long, device=device)
text_token_len = torch.tensor([text_token.shape[1]], dtype=torch.long, device=device) text_token_len = torch.tensor([text_token.shape[1]], dtype=torch.long, device=device)
cur_token, feat = self.tts_model.streaming_one_step( cur_token, feat = self.tts_model.streaming_one_step(
@ -3145,6 +3147,7 @@ class LLMASRXvecSlotTTS(nn.Module):
outside_prompt_lengths=llm_cur_kv_cache_len, outside_prompt_lengths=llm_cur_kv_cache_len,
sampling="threshold_1e-6", sampling="threshold_1e-6",
chunk_idx=chunk_idx, chunk_idx=chunk_idx,
diff_steps=5,
) )
if cur_token is not None and cur_token.shape[1] > 0 and feat.shape[2] > 0: if cur_token is not None and cur_token.shape[1] > 0 and feat.shape[2] > 0:
# process first package, token in B,T,D, feat in B,F,T # process first package, token in B,T,D, feat in B,F,T
@ -3164,8 +3167,6 @@ class LLMASRXvecSlotTTS(nn.Module):
prompt_audio[0] = torch.concat([prompt_audio[0], feat.transpose(1, 2)], dim=1) prompt_audio[0] = torch.concat([prompt_audio[0], feat.transpose(1, 2)], dim=1)
wav = self.vocoder.inference(feat.transpose(1, 2)) wav = self.vocoder.inference(feat.transpose(1, 2))
chunk_idx += 1 chunk_idx += 1
else:
cur_token, feat, wav = None, None, None
return ((cur_token, feat, wav), (text, last_t_size, prompt_token, prompt_audio, chunk_idx)) return ((cur_token, feat, wav), (text, last_t_size, prompt_token, prompt_audio, chunk_idx))