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

Merge branch 'dev_xw' of github.com:alibaba-damo-academy/FunASR into dev_xw

Browse Source 
add
This commit is contained in:
游雁 2023-03-15 20:38:48 +08:00
commit 5b355e0f93
1 changed files with 39 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
View File

@ -23,6 +23,8 @@ class Paraformer():
def __init__(self, model_dir: Union[str, Path] = None,
batch_size: int = 1,
device_id: Union[str, int] = "-1",
plot_timestamp_to: str = "",
pred_bias: int = 1,
):
if not Path(model_dir).exists():
@ -41,14 +43,15 @@ class Paraformer():
)
self.ort_infer = OrtInferSession(model_file, device_id)
self.batch_size = batch_size
self.plot = True
self.plot_timestamp_to = plot_timestamp_to
self.pred_bias = pred_bias
def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
waveform_nums = len(waveform_list)
asr_res = []
for beg_idx in range(0, waveform_nums, self.batch_size):
res = {}
end_idx = min(waveform_nums, beg_idx + self.batch_size)
feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
try:
@ -64,19 +67,41 @@ class Paraformer():
logging.warning("input wav is silence or noise")
preds = ['']
else:
preds, raw_token = self.decode(am_scores, valid_token_lens)[0]
res['preds'] = preds
if us_cif_peak is not None:
timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak, copy.copy(raw_token))
res['timestamp'] = timestamp
if self.plot:
self.plot_wave_timestamp(waveform_list[0], timestamp_total)
asr_res.append(res)
preds = self.decode(am_scores, valid_token_lens)
if us_cif_peak is None:
for pred in preds:
asr_res.append({'preds': pred})
else:
for pred, us_cif_peak_ in zip(preds, us_cif_peak):
text, tokens = pred
timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak_, copy.copy(tokens))
if len(self.plot_timestamp_to):
self.plot_wave_timestamp(waveform_list[0], timestamp_total, self.plot_timestamp_to)
asr_res.append({'preds': text, 'timestamp': timestamp})
return asr_res
def plot_wave_timestamp(self, wav, text_timestamp):
def plot_wave_timestamp(self, wav, text_timestamp, dest):
# TODO: Plot the wav and timestamp results with matplotlib
import pdb; pdb.set_trace()
import matplotlib
matplotlib.use('Agg')
matplotlib.rc("font", family='Alibaba PuHuiTi') # set it to a font that your system supports
import matplotlib.pyplot as plt
fig, ax1 = plt.subplots(figsize=(11, 3.5), dpi=320)
ax2 = ax1.twinx()
ax2.set_ylim([0, 2.0])
# plot waveform
ax1.set_ylim([-0.3, 0.3])
time = np.arange(wav.shape[0]) / 16000
ax1.plot(time, wav/wav.max()*0.3, color='gray', alpha=0.4)
# plot lines and text
for (char, start, end) in text_timestamp:
ax1.vlines(start, -0.3, 0.3, ls='--')
ax1.vlines(end, -0.3, 0.3, ls='--')
x_adj = 0.045 if char != '<sil>' else 0.12
ax1.text((start + end) * 0.5 - x_adj, 0, char)
# plt.legend()
plotname = "{}/timestamp.png".format(dest)
plt.savefig(plotname, bbox_inches='tight')
def load_data(self,
wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:
@ -150,9 +175,7 @@ class Paraformer():
# Change integer-ids to tokens
token = self.converter.ids2tokens(token_int)
# token = token[:valid_token_num-1]
token = token[:valid_token_num-self.pred_bias]
texts = sentence_postprocess(token)
text = texts[0]
# text = self.tokenizer.tokens2text(token)
return text, token
return texts