Added unconditional generation
Browse files
app.py
CHANGED
@@ -180,29 +180,49 @@ def make_npz_prompt(name, uploaded_audio, recorded_audio, transcript_content):
|
|
180 |
def infer_from_audio(text, language, accent, audio_prompt, record_audio_prompt, transcript_content):
|
181 |
if len(text) > 150:
|
182 |
return "Rejected, Text too long (should be less than 150 characters)", None
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
if not isinstance(wav_pr, torch.FloatTensor):
|
188 |
-
wav_pr = torch.FloatTensor(wav_pr)
|
189 |
-
if wav_pr.abs().max() > 1:
|
190 |
-
wav_pr /= wav_pr.abs().max()
|
191 |
-
if wav_pr.size(-1) == 2:
|
192 |
-
wav_pr = wav_pr[:, 0]
|
193 |
-
if wav_pr.ndim == 1:
|
194 |
-
wav_pr = wav_pr.unsqueeze(0)
|
195 |
-
assert wav_pr.ndim and wav_pr.size(0) == 1
|
196 |
-
|
197 |
-
if transcript_content == "":
|
198 |
-
lang_pr, text_pr = transcribe_one(wav_pr, sr)
|
199 |
-
lang_token = lang2token[lang_pr]
|
200 |
-
text_pr = lang_token + text_pr + lang_token
|
201 |
else:
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
206 |
|
207 |
if language == 'auto-detect':
|
208 |
lang_token = lang2token[langid.classify(text)[0]]
|
@@ -212,13 +232,6 @@ def infer_from_audio(text, language, accent, audio_prompt, record_audio_prompt,
|
|
212 |
text = text.replace("\n", "")
|
213 |
text = lang_token + text + lang_token
|
214 |
|
215 |
-
if lang_pr not in ['ja', 'zh', 'en']:
|
216 |
-
return f"Reference audio must be a speech of one of model-supported languages, got {lang_pr} instead", None
|
217 |
-
|
218 |
-
# tokenize audio
|
219 |
-
encoded_frames = tokenize_audio(audio_tokenizer, (wav_pr, sr))
|
220 |
-
audio_prompts = encoded_frames[0][0].transpose(2, 1).to(device)
|
221 |
-
|
222 |
# tokenize text
|
223 |
logging.info(f"synthesize text: {text}")
|
224 |
phone_tokens, langs = text_tokenizer.tokenize(text=f"_{text}".strip())
|
@@ -228,14 +241,7 @@ def infer_from_audio(text, language, accent, audio_prompt, record_audio_prompt,
|
|
228 |
]
|
229 |
)
|
230 |
|
231 |
-
|
232 |
-
if text_pr:
|
233 |
-
text_prompts, _ = text_tokenizer.tokenize(text=f"{text_pr}".strip())
|
234 |
-
text_prompts, enroll_x_lens = text_collater(
|
235 |
-
[
|
236 |
-
text_prompts
|
237 |
-
]
|
238 |
-
)
|
239 |
text_tokens = torch.cat([text_prompts, text_tokens], dim=-1)
|
240 |
text_tokens_lens += enroll_x_lens
|
241 |
lang = lang if accent == "no-accent" else token2lang[langdropdown2token[accent]]
|
|
|
180 |
def infer_from_audio(text, language, accent, audio_prompt, record_audio_prompt, transcript_content):
|
181 |
if len(text) > 150:
|
182 |
return "Rejected, Text too long (should be less than 150 characters)", None
|
183 |
+
if audio_prompt is None and record_audio_prompt is None:
|
184 |
+
audio_prompts = torch.zeros([1, 0, NUM_QUANTIZERS]).type(torch.int32).to(device)
|
185 |
+
text_prompts = torch.zeros([1, 0]).type(torch.int32)
|
186 |
+
lang_pr = language if language != 'mix' else 'en'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
187 |
else:
|
188 |
+
audio_prompt = audio_prompt if audio_prompt is not None else record_audio_prompt
|
189 |
+
sr, wav_pr = audio_prompt
|
190 |
+
if len(wav_pr) / sr > 15:
|
191 |
+
return "Rejected, Audio too long (should be less than 15 seconds)", None
|
192 |
+
if not isinstance(wav_pr, torch.FloatTensor):
|
193 |
+
wav_pr = torch.FloatTensor(wav_pr)
|
194 |
+
if wav_pr.abs().max() > 1:
|
195 |
+
wav_pr /= wav_pr.abs().max()
|
196 |
+
if wav_pr.size(-1) == 2:
|
197 |
+
wav_pr = wav_pr[:, 0]
|
198 |
+
if wav_pr.ndim == 1:
|
199 |
+
wav_pr = wav_pr.unsqueeze(0)
|
200 |
+
assert wav_pr.ndim and wav_pr.size(0) == 1
|
201 |
+
|
202 |
+
if transcript_content == "":
|
203 |
+
lang_pr, text_pr = transcribe_one(wav_pr, sr)
|
204 |
+
lang_token = lang2token[lang_pr]
|
205 |
+
text_pr = lang_token + text_pr + lang_token
|
206 |
+
else:
|
207 |
+
lang_pr = langid.classify(str(transcript_content))[0]
|
208 |
+
text_pr = transcript_content.replace("\n", "")
|
209 |
+
if lang_pr not in ['ja', 'zh', 'en']:
|
210 |
+
return f"Reference audio must be a speech of one of model-supported languages, got {lang_pr} instead", None
|
211 |
+
lang_token = lang2token[lang_pr]
|
212 |
+
text_pr = lang_token + text_pr + lang_token
|
213 |
+
|
214 |
+
# tokenize audio
|
215 |
+
encoded_frames = tokenize_audio(audio_tokenizer, (wav_pr, sr))
|
216 |
+
audio_prompts = encoded_frames[0][0].transpose(2, 1).to(device)
|
217 |
+
|
218 |
+
enroll_x_lens = None
|
219 |
+
if text_pr:
|
220 |
+
text_prompts, _ = text_tokenizer.tokenize(text=f"{text_pr}".strip())
|
221 |
+
text_prompts, enroll_x_lens = text_collater(
|
222 |
+
[
|
223 |
+
text_prompts
|
224 |
+
]
|
225 |
+
)
|
226 |
|
227 |
if language == 'auto-detect':
|
228 |
lang_token = lang2token[langid.classify(text)[0]]
|
|
|
232 |
text = text.replace("\n", "")
|
233 |
text = lang_token + text + lang_token
|
234 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
# tokenize text
|
236 |
logging.info(f"synthesize text: {text}")
|
237 |
phone_tokens, langs = text_tokenizer.tokenize(text=f"_{text}".strip())
|
|
|
241 |
]
|
242 |
)
|
243 |
|
244 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
text_tokens = torch.cat([text_prompts, text_tokens], dim=-1)
|
246 |
text_tokens_lens += enroll_x_lens
|
247 |
lang = lang if accent == "no-accent" else token2lang[langdropdown2token[accent]]
|