msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

Merge 4be4292658 into 94987f26df

Browse Source
pull/3096/merge
i4never 2 years ago committed by GitHub
parent 94987f26df 4be4292658
commit a726102c3e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 118 additions and 5 deletions
Show Stats Download Patch File Download Diff File

40
paddlespeech/cli/tts/infer.py
Unescape View File

@ -457,6 +457,16 @@ class TTSExecutor(BaseExecutor):
text: str, text: str,
lang: str='zh', lang: str='zh',
am: str='fastspeech2_csmsc', am: str='fastspeech2_csmsc',
durations_scale: float=None,
durations_bias: float=None,
pitch_scale: float = None,
pitch_bias: float = None,
pitch_stats_mean: float = None,
pitch_stats_std: float = None,
energy_scale: float = None,
energy_bias: float = None,
energy_stats_mean: float = None,
energy_stats_std: float = None,
spk_id: int=0): spk_id: int=0):
""" """
Model inference and result stored in self.output. Model inference and result stored in self.output.
@ -493,7 +503,16 @@ class TTSExecutor(BaseExecutor):
mel = self.am_inference( mel = self.am_inference(
part_phone_ids, spk_id=paddle.to_tensor([spk_id])) part_phone_ids, spk_id=paddle.to_tensor([spk_id]))
else: else:
mel = self.am_inference(part_phone_ids) use_teacher_forcing = any([a is not None for a in [durations_scale, durations_bias,
pitch_scale, pitch_bias,
energy_scale, energy_bias]])
mel = self.am_inference(part_phone_ids,
durations_scale=durations_scale, durations_bias=durations_bias,
pitch_scale=pitch_scale, pitch_bias=pitch_bias,
pitch_stats_mean=pitch_stats_mean, pitch_stats_std=pitch_stats_std,
energy_scale=energy_scale, energy_bias=energy_bias,
energy_stats_mean=energy_stats_mean, energy_stats_std=energy_stats_std,
use_teacher_forcing=use_teacher_forcing)
self.am_time += (time.time() - am_st) self.am_time += (time.time() - am_st)
# voc # voc
voc_st = time.time() voc_st = time.time()
@ -687,7 +706,17 @@ class TTSExecutor(BaseExecutor):
output: str='output.wav', output: str='output.wav',
use_onnx: bool=False, use_onnx: bool=False,
cpu_threads: int=2, cpu_threads: int=2,
fs: int=24000): fs: int=24000,
durations_scale=None,
durations_bias=None,
pitch_scale=None,
pitch_bias=None,
pitch_stats_mean=None,
pitch_stats_std=None,
energy_scale=None,
energy_bias=None,
energy_stats_mean=None,
energy_stats_std=None):
""" """
Python API to call an executor. Python API to call an executor.
""" """
@ -707,7 +736,12 @@ class TTSExecutor(BaseExecutor):
voc_stat=voc_stat, voc_stat=voc_stat,
lang=lang) lang=lang)
self.infer(text=text, lang=lang, am=am, spk_id=spk_id) self.infer(text=text, lang=lang, am=am, spk_id=spk_id,
durations_scale=durations_scale, durations_bias=durations_bias,
pitch_scale=pitch_scale, pitch_bias=pitch_bias,
pitch_stats_mean=pitch_stats_mean, pitch_stats_std=pitch_stats_std,
energy_scale=energy_scale, energy_bias=energy_bias,
energy_stats_mean=energy_stats_mean, energy_stats_std=energy_stats_std)
res = self.postprocess(output=output) res = self.postprocess(output=output)
return res return res
else: else:

83
paddlespeech/t2s/models/fastspeech2/fastspeech2.py
Unescape View File

@ -596,6 +596,16 @@ class FastSpeech2(nn.Layer):
ds: paddle.Tensor=None, ds: paddle.Tensor=None,
ps: paddle.Tensor=None, ps: paddle.Tensor=None,
es: paddle.Tensor=None, es: paddle.Tensor=None,
durations_scale: float=None,
durations_bias: float=None,
pitch_scale: float = None,
pitch_bias: float = None,
pitch_stats_mean: float = None,
pitch_stats_std: float = None,
energy_scale: float = None,
energy_bias: float = None,
energy_stats_mean: float = None,
energy_stats_std: float = None,
is_inference: bool=False, is_inference: bool=False,
return_after_enc=False, return_after_enc=False,
alpha: float=1.0, alpha: float=1.0,
@ -645,10 +655,32 @@ class FastSpeech2(nn.Layer):
d_outs = ds d_outs = ds
else: else:
d_outs = self.duration_predictor.inference(hs, d_masks) d_outs = self.duration_predictor.inference(hs, d_masks)
if durations_scale is not None or durations_bias is not None:
durations_scale = durations_scale if durations_scale is not None else 1
durations_bias = durations_bias if durations_bias is not None else 0
d_outs = durations_scale * d_outs + durations_bias
# pitch控制
if ps is not None: if ps is not None:
p_outs = ps p_outs = ps
elif pitch_scale is not None or pitch_bias is not None:
pitch_scale = pitch_scale if pitch_scale is not None else 1
pitch_bias = pitch_bias if pitch_bias is not None else 0
assert pitch_stats_mean is not None and pitch_stats_std is not None
p_outs = paddle.exp(p_outs * pitch_stats_std + pitch_stats_mean)
p_outs = p_outs * pitch_scale + pitch_bias
p_outs = (paddle.log(p_outs) - pitch_stats_mean) / pitch_stats_std
# energy控制
if es is not None: if es is not None:
e_outs = es e_outs = es
elif energy_scale is not None or energy_bias is not None:
energy_scale = energy_scale if energy_scale is not None else 1
energy_bias = energy_bias if energy_bias is not None else 0
assert energy_stats_mean is not None and energy_stats_std is not None
e_outs = paddle.exp(e_outs * energy_stats_std + energy_stats_mean)
e_outs = e_outs * energy_scale + energy_bias
e_outs = (paddle.log(e_outs) - energy_stats_mean) / energy_stats_std
# use prediction in inference # use prediction in inference
# (B, Tmax, 1) # (B, Tmax, 1)
@ -747,6 +779,16 @@ class FastSpeech2(nn.Layer):
durations: paddle.Tensor=None, durations: paddle.Tensor=None,
pitch: paddle.Tensor=None, pitch: paddle.Tensor=None,
energy: paddle.Tensor=None, energy: paddle.Tensor=None,
durations_scale: float = None,
durations_bias: float = None,
pitch_scale: float = None,
pitch_bias: float = None,
pitch_stats_mean: float = None,
pitch_stats_std: float = None,
energy_scale: float = None,
energy_bias: float = None,
energy_stats_mean: float = None,
energy_stats_std: float = None,
alpha: float=1.0, alpha: float=1.0,
use_teacher_forcing: bool=False, use_teacher_forcing: bool=False,
spk_emb=None, spk_emb=None,
@ -764,6 +806,26 @@ class FastSpeech2(nn.Layer):
Groundtruth of token-averaged pitch (T, 1). Groundtruth of token-averaged pitch (T, 1).
energy(Tensor, optional): energy(Tensor, optional):
Groundtruth of token-averaged energy (T, 1). Groundtruth of token-averaged energy (T, 1).
durations_scale: (int, float, optional):
For duration control during infer
durations_bias: (int, float, optional):
For duration control during infer
pitch_scale:(int, float, optional):
For pitch control during infer
pitch_bias:(int, float, optional):
For pitch control during infer
pitch_stats_mean:(int, float, optional):
For pitch control during infer
pitch_stats_std:(int, float, optional):
For pitch control during infer
energy_scale:(int, float, optional):
For energy control during infer
energy_bias:(int, float, optional):
For energy control during infer
energy_stats_mean:(int, float, optional):
For energy control during infer
energy_stats_std:(int, float, optional):
For energy control during infer
alpha(float, optional): alpha(float, optional):
Alpha to control the speed. Alpha to control the speed.
use_teacher_forcing(bool, optional): use_teacher_forcing(bool, optional):
@ -806,9 +868,20 @@ class FastSpeech2(nn.Layer):
ds=ds, ds=ds,
ps=ps, ps=ps,
es=es, es=es,
durations_scale=durations_scale,
durations_bias=durations_bias,
pitch_scale=pitch_scale,
pitch_bias=pitch_bias,
pitch_stats_mean=pitch_stats_mean,
pitch_stats_std=pitch_stats_std,
energy_scale=energy_scale,
energy_bias=energy_bias,
energy_stats_mean=energy_stats_mean,
energy_stats_std=energy_stats_std,
spk_emb=spk_emb, spk_emb=spk_emb,
spk_id=spk_id, spk_id=spk_id,
tone_id=tone_id, tone_id=tone_id,
alpha=alpha,
is_inference=True) is_inference=True)
else: else:
# (1, L, odim) # (1, L, odim)
@ -921,9 +994,15 @@ class FastSpeech2Inference(nn.Layer):
self.normalizer = normalizer self.normalizer = normalizer
self.acoustic_model = model self.acoustic_model = model
def forward(self, text, spk_id=None, spk_emb=None): def forward(self, text, spk_id=None, spk_emb=None, durations_scale=None, durations_bias=None, pitch_scale=None,
pitch_bias=None, pitch_stats_mean=None, pitch_stats_std=None, energy_scale=None, energy_bias=None,
energy_stats_mean=None, energy_stats_std=None, use_teacher_forcing=False):
normalized_mel, d_outs, p_outs, e_outs = self.acoustic_model.inference( normalized_mel, d_outs, p_outs, e_outs = self.acoustic_model.inference(
text, spk_id=spk_id, spk_emb=spk_emb) text, spk_id=spk_id, spk_emb=spk_emb, durations_scale=durations_scale, durations_bias=durations_bias,
pitch_scale=pitch_scale, pitch_bias=pitch_bias, pitch_stats_mean=pitch_stats_mean,
pitch_stats_std=pitch_stats_std, energy_scale=energy_scale, energy_bias=energy_bias,
energy_stats_mean=energy_stats_mean, energy_stats_std=energy_stats_std,
use_teacher_forcing=use_teacher_forcing)
logmel = self.normalizer.inverse(normalized_mel) logmel = self.normalizer.inverse(normalized_mel)
return logmel return logmel

Write Preview
Loading…
Cancel
Save