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
Funasr1.0 (#1275)
zhifu gao2cca8104d26b
diff --git a/examples/industrial_data_pretraining/paraformer/finetune.sh b/examples/industrial_data_pretraining/paraformer/finetune.shindex 7d8987602..1aff0683c 100644--- a/examples/industrial_data_pretraining/paraformer/finetune.sh+++ b/examples/industrial_data_pretraining/paraformer/finetune.sh@@ -11,9 +11,9 @@ python funasr/bin/train.py \+model_revision="v2.0.2" \+train_data_set_list="/Users/zhifu/funasr_github/test_local/aishell2_dev_ios/asr_task_debug_len_10.jsonl" \+valid_data_set_list="/Users/zhifu/funasr_github/test_local/aishell2_dev_ios/asr_task_debug_len_10.jsonl" \-++dataset_conf.batch_size=2 \+++dataset_conf.batch_size=64 \++dataset_conf.batch_type="example" \++train_conf.max_epoch=2 \+++dataset_conf.num_workers=4 \+output_dir="outputs/debug/ckpt/funasr2/exp2" \-+device="cpu" \+debug="true"\ No newline at end of filediff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex bedc17d16..332013612 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -132,7 +132,7 @@ class AutoModel:self.punc_kwargs = punc_kwargsself.spk_model = spk_modelself.spk_kwargs = spk_kwargs- self.model_path = kwargs["model_path"]+ self.model_path = kwargs.get("model_path", "./")def build_model(self, **kwargs):diff --git a/funasr/bin/train.py b/funasr/bin/train.pyindex 0334006c5..d9d4d6241 100644--- a/funasr/bin/train.py+++ b/funasr/bin/train.py@@ -40,7 +40,7 @@ def main_hydra(kwargs: DictConfig):def main(**kwargs):-+ print(kwargs)# set random seedtables.print()set_all_random_seed(kwargs.get("seed", 0))diff --git a/funasr/datasets/audio_datasets/samplers.py b/funasr/datasets/audio_datasets/samplers.pyindex e170c681b..0d9309814 100644--- a/funasr/datasets/audio_datasets/samplers.py+++ b/funasr/datasets/audio_datasets/samplers.py@@ -28,7 +28,7 @@ class BatchSampler(torch.utils.data.BatchSampler):self.shuffle = shuffle and is_trainingdef __len__(self):- return self.total_samples+ return (self.total_samples-1) // self.batch_size + 1def set_epoch(self, epoch):np.random.seed(epoch)diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.pyindex 7c2156174..becfd56e3 100644--- a/funasr/models/fsmn_vad_streaming/model.py+++ b/funasr/models/fsmn_vad_streaming/model.py@@ -255,7 +255,6 @@ class Stats(object):self.waveform = Noneself.last_drop_frames = 0-@tables.register("model_classes", "FsmnVADStreaming")class FsmnVADStreaming(nn.Module):"""@@ -500,7 +499,6 @@ class FsmnVADStreaming(nn.Module):# # reset class variables and clear the dict for the next query# self.AllResetDetection()return segments-def init_cache(self, cache: dict = {}, **kwargs):diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.pyindex 91b30b0a8..62d6be80b 100644--- a/funasr/train_utils/trainer.py+++ b/funasr/train_utils/trainer.py@@ -147,9 +147,17 @@ class Trainer:for epoch in range(self.start_epoch, self.max_epoch + 1):self._train_epoch(epoch)++ if self.use_ddp or self.use_fsdp:+ dist.barrier()+self._validate_epoch(epoch)-++ if self.use_ddp or self.use_fsdp:+ dist.barrier()++if self.rank == 0:self._save_checkpoint(epoch)@@ -164,7 +172,9 @@ class Trainer:if self.use_ddp or self.use_fsdp:dist.barrier()- self.writer.close()++ if self.writer:+ self.writer.close()def _train_epoch(self, epoch):@@ -230,6 +240,8 @@ class Trainer:continue# Execute an optimization step (update model parameters)+ if self.use_ddp or self.use_fsdp:+ dist.barrier()self.optim.step()self.scheduler.step()# Clear gradients for the next accumulation stage@@ -244,7 +256,7 @@ class Trainer:pbar.update(1)if self.local_rank == 0:description = (- f"Epoch: {epoch}/{self.max_epoch}, "+ f"Train epoch: {epoch}/{self.max_epoch}, "f"step {batch_idx}/{len(self.dataloader_train)}, "f"{speed_stats}, "f"(loss: {loss.detach().cpu().item():.3f}), "@@ -306,7 +318,7 @@ class Trainer:pbar.update(1)if self.local_rank == 0:description = (- f"validation: \nEpoch: {epoch}/{self.max_epoch}, "+ f"validation epoch: {epoch}/{self.max_epoch}, "f"step {batch_idx}/{len(self.dataloader_train)}, "f"{speed_stats}, "f"(loss: {loss.detach().cpu().item():.3f}), "