Skip to content

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

decoding

游雁45d7aa900476
diff --cc docs/images/wechat.png
index 705ff75de,491a1c3df..8d37700c9
Binary files differ
diff --cc examples/industrial_data_pretraining/ctc/demo.py
index 85a748a65,1afb4b54f..b9d1647b7
--- a/examples/industrial_data_pretraining/ctc/demo.py
+++ b/examples/industrial_data_pretraining/ctc/demo.py
@@@ -6,8 -6,10 +6,11 @@@
import sys
from funasr import AutoModel
- model_dir=sys.argv[1]
- input_file=sys.argv[2]
++
+ model_dir = "/Users/zhifu/Downloads/modelscope_models/ctc_model"
+ input_file = (
+ "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
+ )
model = AutoModel(
model=model_dir,
diff --cc funasr/auto/auto_model.py
index 91e80d8be,e30eb0976..a6cd3a652
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@@ -324,7 -324,7 +324,7 @@@ class AutoModel
input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg
)
end_vad = time.time()
--
++
# FIX(gcf): concat the vad clips for sense vocie model for better aed
if kwargs.get("merge_vad", False):
for i in range(len(res)):
@@@ -466,22 -466,25 +466,22 @@@
result[k] = restored_data[j][k]
else:
result[k] += restored_data[j][k]
-
+
+ if not len(result["text"].strip()):
+ continue
return_raw_text = kwargs.get("return_raw_text", False)
# step.3 compute punc model
+ raw_text = None
if self.punc_model is not None:
- if not len(result["text"].strip()):
- if return_raw_text:
- result["raw_text"] = ""
- else:
- deep_update(self.punc_kwargs, cfg)
- punc_res = self.inference(
- result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg
- )
- raw_text = copy.copy(result["text"])
- if return_raw_text:
- result["raw_text"] = raw_text
- result["text"] = punc_res[0]["text"]
- else:
- raw_text = None
+ deep_update(self.punc_kwargs, cfg)
+ punc_res = self.inference(
+ result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg
+ )
+ raw_text = copy.copy(result["text"])
+ if return_raw_text:
+ result["raw_text"] = raw_text
+ result["text"] = punc_res[0]["text"]
-
+
# speaker embedding cluster after resorted
if self.spk_model is not None and kwargs.get("return_spk_res", True):
if raw_text is None:
@@@ -602,6 -605,12 +602,6 @@@
)
with torch.no_grad():
- export_dir = export_utils.export(model=model, data_in=data_list, **kwargs)
-
- if type == "onnx":
- export_dir = export_utils.export_onnx(model=model, data_in=data_list, **kwargs)
- else:
- export_dir = export_utils.export_torchscripts(
- model=model, data_in=data_list, **kwargs
- )
++ export_dir = export_utils.export(model=model, data_in=data_list, **kwargs)
return export_dir
diff --cc funasr/models/paraformer/cif_predictor.py
index 05e283a24,7490310c9..0856eede2
--- a/funasr/models/paraformer/cif_predictor.py
+++ b/funasr/models/paraformer/cif_predictor.py
@@@ -494,6 -494,6 +494,8 @@@ class CifPredictorV2Export(torch.nn.Mod
token_num_floor = torch.floor(token_num)
return hidden, alphas, token_num_floor
++
++
@torch.jit.script
def cif_v1_export(hidden, alphas, threshold: float):
device = hidden.device
@@@ -516,9 -516,9 +518,7 @@@
fires[fire_idxs] = 1
fires = fires + prefix_sum - prefix_sum_floor
-- prefix_sum_hidden = torch.cumsum(
-- alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1
-- )
++ prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
frames = prefix_sum_hidden[fire_idxs]
shift_frames = torch.roll(frames, 1, dims=0)
@@@ -530,26 -530,25 +530,22 @@@
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
-- remain_frames = (
-- remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
-- )
++ remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
shift_remain_frames[shift_batch_idxs] = 0
frames = frames - shift_frames + shift_remain_frames - remain_frames
- max_label_len = alphas.sum(dim=-1)
- max_label_len = torch.floor(max_label_len).max().to(dtype=torch.int64)
+ max_label_len = batch_len.max()
-- frame_fires = torch.zeros(
-- batch_size, max_label_len, hidden_size, dtype=dtype, device=device
-- )
++ frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)
frame_fires[frame_fires_idxs] = frames
return frame_fires, fires
++
@torch.jit.script
def cif_export(hidden, alphas, threshold: float):
batch_size, len_time, hidden_size = hidden.size()
@@@ -693,11 -692,11 +689,8 @@@ def cif_v1(hidden, alphas, threshold)
device = hidden.device
dtype = hidden.dtype
batch_size, len_time, hidden_size = hidden.size()
-- frames = torch.zeros(batch_size, len_time, hidden_size,
-- dtype=dtype, device=device)
-- prefix_sum_hidden = torch.cumsum(
-- alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1
-- )
++ frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
++ prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
frames = prefix_sum_hidden[fire_idxs]
shift_frames = torch.roll(frames, 1, dims=0)
@@@ -709,21 -708,21 +702,16 @@@
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
-- remain_frames = (
-- remains[fire_idxs].unsqueeze(-1).tile((1,
-- hidden_size)) * hidden[fire_idxs]
-- )
++ remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
shift_remain_frames[shift_batch_idxs] = 0
frames = frames - shift_frames + shift_remain_frames - remain_frames
- max_label_len = torch.round(alphas.sum(-1)).int().max() # torch.round to calculate the max length
+ max_label_len = batch_len.max()
-- frame_fires = torch.zeros(
-- batch_size, max_label_len, hidden_size, dtype=dtype, device=device
-- )
++ frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)
frame_fires[frame_fires_idxs] = frames
diff --cc funasr/models/sense_voice/model.py
index 97f1b1913,697f50c00..a9b2149af
--- a/funasr/models/sense_voice/model.py
+++ b/funasr/models/sense_voice/model.py
@@@ -1384,3 -1394,269 +1386,279 @@@ class SenseVoiceSANM(nn.Module)
ibest_writer["text"][key[i]] = text
return results, meta_data
+
+
+ from funasr.models.paraformer.search import Hypothesis
+ from funasr.utils import postprocess_utils
+
+
+ @tables.register("model_classes", "SenseVoiceSANMCTC")
+ class SenseVoiceSANMCTC(nn.Module):
+ """CTC-attention hybrid Encoder-Decoder model"""
+
+ def __init__(
+ self,
+ specaug: str = None,
+ specaug_conf: dict = None,
+ normalize: str = None,
+ normalize_conf: dict = None,
+ encoder: str = None,
+ encoder_conf: dict = None,
+ ctc_conf: dict = None,
+ input_size: int = 80,
+ vocab_size: int = -1,
+ ignore_id: int = -1,
+ blank_id: int = 0,
+ sos: int = 1,
+ eos: int = 2,
+ length_normalized_loss: bool = False,
+ **kwargs,
+ ):
+
+ super().__init__()
+
+ if specaug is not None:
+ specaug_class = tables.specaug_classes.get(specaug)
+ specaug = specaug_class(**specaug_conf)
+ if normalize is not None:
+ normalize_class = tables.normalize_classes.get(normalize)
+ normalize = normalize_class(**normalize_conf)
+ encoder_class = tables.encoder_classes.get(encoder)
+ encoder = encoder_class(input_size=input_size, **encoder_conf)
+ encoder_output_size = encoder.output_size()
+
+ if ctc_conf is None:
+ ctc_conf = {}
+ ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)
+
+ self.blank_id = blank_id
+ self.sos = sos if sos is not None else vocab_size - 1
+ self.eos = eos if eos is not None else vocab_size - 1
+ self.vocab_size = vocab_size
+ self.ignore_id = ignore_id
+ self.specaug = specaug
+ self.normalize = normalize
+ self.encoder = encoder
+ self.error_calculator = None
+
+ self.ctc = ctc
+
+ self.length_normalized_loss = length_normalized_loss
+ self.encoder_output_size = encoder_output_size
+
+ self.lid_dict = {"zh": 3, "en": 4, "yue": 7, "ja": 11, "ko": 12, "nospeech": 13}
+ self.textnorm_dict = {"withtextnorm": 14, "wotextnorm": 15}
+ self.embed = torch.nn.Embedding(8 + len(self.lid_dict) + len(self.textnorm_dict), 560)
+
+ def forward(
+ self,
+ speech: torch.Tensor,
+ speech_lengths: torch.Tensor,
+ text: torch.Tensor,
+ text_lengths: torch.Tensor,
+ **kwargs,
+ ):
+ """Encoder + Decoder + Calc loss
+ Args:
+ speech: (Batch, Length, ...)
+ speech_lengths: (Batch, )
+ text: (Batch, Length)
+ text_lengths: (Batch,)
+ """
+ # import pdb;
+ # pdb.set_trace()
+ if len(text_lengths.size()) > 1:
+ text_lengths = text_lengths[:, 0]
+ if len(speech_lengths.size()) > 1:
+ speech_lengths = speech_lengths[:, 0]
+
+ batch_size = speech.shape[0]
+
+ # 1. Encoder
+ encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+
+ loss_ctc, cer_ctc = None, None
+ stats = dict()
+
+ loss_ctc, cer_ctc = self._calc_ctc_loss(encoder_out, encoder_out_lens, text, text_lengths)
+
+ loss = loss_ctc
+
+ # Collect total loss stats
+ stats["loss"] = torch.clone(loss.detach())
+
+ # force_gatherable: to-device and to-tensor if scalar for DataParallel
+ if self.length_normalized_loss:
+ batch_size = int((text_lengths + 1).sum())
+ loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
+ return loss, stats, weight
+
+ def encode(
+ self,
+ speech: torch.Tensor,
+ speech_lengths: torch.Tensor,
+ **kwargs,
+ ):
+ """Frontend + Encoder. Note that this method is used by asr_inference.py
+ Args:
+ speech: (Batch, Length, ...)
+ speech_lengths: (Batch, )
+ ind: int
+ """
+
+ # Data augmentation
+ if self.specaug is not None and self.training:
+ speech, speech_lengths = self.specaug(speech, speech_lengths)
+
+ # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
+ if self.normalize is not None:
+ speech, speech_lengths = self.normalize(speech, speech_lengths)
+
+ # Forward encoder
+ # feats: (Batch, Length, Dim)
+ # -> encoder_out: (Batch, Length2, Dim2)
+ encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)
+
+ return encoder_out, encoder_out_lens
+
+ def _calc_ctc_loss(
+ self,
+ encoder_out: torch.Tensor,
+ encoder_out_lens: torch.Tensor,
+ ys_pad: torch.Tensor,
+ ys_pad_lens: torch.Tensor,
+ ):
+ # Calc CTC loss
+ loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
+
+ # Calc CER using CTC
+ cer_ctc = None
+ if not self.training and self.error_calculator is not None:
+ ys_hat = self.ctc.argmax(encoder_out).data
+ cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
+ return loss_ctc, cer_ctc
+
+ def inference(
+ self,
+ data_in,
+ data_lengths=None,
+ key: list = None,
+ tokenizer=None,
+ frontend=None,
+ **kwargs,
+ ):
+
+ if kwargs.get("batch_size", 1) > 1:
+ raise NotImplementedError("batch decoding is not implemented")
+
+ meta_data = {}
+ if (
+ isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
+ ): # fbank
+ speech, speech_lengths = data_in, data_lengths
+ if len(speech.shape) < 3:
+ speech = speech[None, :, :]
+ if speech_lengths is None:
+ speech_lengths = speech.shape[1]
+ else:
+ # extract fbank feats
+ time1 = time.perf_counter()
+ audio_sample_list = load_audio_text_image_video(
+ data_in,
+ fs=frontend.fs,
+ audio_fs=kwargs.get("fs", 16000),
+ data_type=kwargs.get("data_type", "sound"),
+ tokenizer=tokenizer,
+ )
+ time2 = time.perf_counter()
+ meta_data["load_data"] = f"{time2 - time1:0.3f}"
+ speech, speech_lengths = extract_fbank(
+ audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
+ )
+ time3 = time.perf_counter()
+ meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
+ meta_data["batch_data_time"] = (
+ speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
+ )
+
+ speech = speech.to(device=kwargs["device"])
+ speech_lengths = speech_lengths.to(device=kwargs["device"])
+
+ language = kwargs.get("language", None)
+ if language is not None:
- language_query = self.embed(torch.LongTensor([[self.lid_dict[language] if language in self.lid_dict else 0]]).to(speech.device)).repeat(speech.size(0), 1, 1)
++ language_query = self.embed(
++ torch.LongTensor(
++ [[self.lid_dict[language] if language in self.lid_dict else 0]]
++ ).to(speech.device)
++ ).repeat(speech.size(0), 1, 1)
+ else:
- language_query = self.embed(torch.LongTensor([[0]]).to(speech.device)).repeat(speech.size(0), 1, 1)
++ language_query = self.embed(torch.LongTensor([[0]]).to(speech.device)).repeat(
++ speech.size(0), 1, 1
++ )
+ textnorm = kwargs.get("text_norm", "wotextnorm")
- textnorm_query = self.embed(torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)).repeat(speech.size(0), 1, 1)
++ textnorm_query = self.embed(
++ torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)
++ ).repeat(speech.size(0), 1, 1)
+ speech = torch.cat((textnorm_query, speech), dim=1)
+ speech_lengths += 1
+
- event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(speech.size(0), 1, 1)
++ event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
++ speech.size(0), 1, 1
++ )
+ input_query = torch.cat((language_query, event_emo_query), dim=1)
+ speech = torch.cat((input_query, speech), dim=1)
+ speech_lengths += 3
+
+ # Encoder
+ encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+ if isinstance(encoder_out, tuple):
+ encoder_out = encoder_out[0]
+
+ # c. Passed the encoder result and the beam search
+ ctc_logits = self.ctc.log_softmax(encoder_out)
+
+ results = []
+ b, n, d = encoder_out.size()
+ if isinstance(key[0], (list, tuple)):
+ key = key[0]
+ if len(key) < b:
+ key = key * b
+ for i in range(b):
+ x = ctc_logits[i, : encoder_out_lens[i], :]
+ yseq = x.argmax(dim=-1)
+ yseq = torch.unique_consecutive(yseq, dim=-1)
+ yseq = torch.tensor([self.sos] + yseq.tolist() + [self.eos], device=yseq.device)
+ nbest_hyps = [Hypothesis(yseq=yseq)]
+
+ for nbest_idx, hyp in enumerate(nbest_hyps):
+ ibest_writer = None
+ if kwargs.get("output_dir") is not None:
+ if not hasattr(self, "writer"):
+ self.writer = DatadirWriter(kwargs.get("output_dir"))
+ ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]
+
+ # remove sos/eos and get results
+ last_pos = -1
+ if isinstance(hyp.yseq, list):
+ token_int = hyp.yseq[1:last_pos]
+ else:
+ token_int = hyp.yseq[1:last_pos].tolist()
+
+ # remove blank symbol id, which is assumed to be 0
+ token_int = list(
+ filter(
+ lambda x: x != self.eos and x != self.sos and x != self.blank_id, token_int
+ )
+ )
+
+ # Change integer-ids to tokens
+ text = tokenizer.decode(token_int)
+
+ result_i = {"key": key[i], "text": text}
+ results.append(result_i)
+
+ if ibest_writer is not None:
+ ibest_writer["token"][key[i]] = " ".join(token)
+ ibest_writer["text"][key[i]] = text_postprocessed
+
+ return results, meta_data