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
perf(models/FsmnVADStreaming): optimize GetFrameState and PopDataToOutputBuf (#2177)
Truco1a45b647a858
diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.pyindex 03aaca779..bfffca88d 100644--- a/funasr/models/fsmn_vad_streaming/model.py+++ b/funasr/models/fsmn_vad_streaming/model.py@@ -406,7 +406,6 @@ class FsmnVADStreaming(nn.Module):cur_seg = cache["stats"].output_data_buf[-1]if cur_seg.end_ms != start_frm * self.vad_opts.frame_in_ms:print("warning\n")- out_pos = len(cur_seg.buffer) # cur_seg.buff现在没做任何操作data_to_pop = 0if end_point_is_sent_end:data_to_pop = expected_sample_number@@ -420,12 +419,6 @@ class FsmnVADStreaming(nn.Module):expected_sample_number = len(cache["stats"].data_buf)cur_seg.doa = 0- for sample_cpy_out in range(0, data_to_pop):- # cur_seg.buffer[out_pos ++] = data_buf_.back();- out_pos += 1- for sample_cpy_out in range(data_to_pop, expected_sample_number):- # cur_seg.buffer[out_pos++] = data_buf_.back()- out_pos += 1if cur_seg.end_ms != start_frm * self.vad_opts.frame_in_ms:print("Something wrong with the VAD algorithm\n")cache["stats"].data_buf_start_frame += frm_cnt@@ -512,10 +505,17 @@ class FsmnVADStreaming(nn.Module):assert len(cache["stats"].sil_pdf_ids) == self.vad_opts.silence_pdf_numif len(cache["stats"].sil_pdf_ids) > 0:assert len(cache["stats"].scores) == 1 # 只支持batch_size = 1的测试- sil_pdf_scores = [- cache["stats"].scores[0][t][sil_pdf_id] for sil_pdf_id in cache["stats"].sil_pdf_ids- ]- sum_score = sum(sil_pdf_scores)+ """+ - Change type of `sum_score` to float. The reason is that `sum_score` is a tensor with single element.+ and `torch.Tensor` is slower `float` when tensor has only one element.+ - Put the iteration of `sil_pdf_ids` inside `sum()` to reduce the overhead of creating a new list.+ - The default `sil_pdf_ids` is [0], the `if` statement is used to reduce the overhead of expression+ generation, which result in a mere (~2%) performance gain.+ """+ if len(cache["stats"].sil_pdf_ids) > 1:+ sum_score = sum(cache["stats"].scores[0][t][sil_pdf_id].item() for sil_pdf_id in cache["stats"].sil_pdf_ids)+ else:+ sum_score = cache["stats"].scores[0][t][cache["stats"].sil_pdf_ids[0]].item()noise_prob = math.log(sum_score) * self.vad_opts.speech_2_noise_ratiototal_score = 1.0sum_score = total_score - sum_score