Sumedh commited on
Commit
22b006c
·
1 Parent(s): cb89656

Claude: fix indic-parler-tts description tokenizer loading + pin parler-tts to Dec 2024 commit

Browse files

- Load description tokenizer from model.config.text_encoder._name_or_path
(per official AI4Bharat example) instead of a nonexistent
subfolder='description_tokenizer', which caused 'Unrecognized model'
- Pass attention masks to generate() for cleaner output
- Pin parler-tts to commit d108732 (Dec 2024) to match indic-parler-tts

Files changed (2) hide show
  1. app.py +12 -4
  2. requirements.txt +1 -1
app.py CHANGED
@@ -121,9 +121,12 @@ def _get_indic_tts():
121
  _indic_tts_model = ParlerTTSForConditionalGeneration.from_pretrained(
122
  "ai4bharat/indic-parler-tts"
123
  ).to(device)
 
124
  _indic_tts_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
 
 
125
  _indic_desc_tokenizer = AutoTokenizer.from_pretrained(
126
- "ai4bharat/indic-parler-tts", subfolder="description_tokenizer"
127
  )
128
  print(f"[IndicTTS] Model ready on {device}.", flush=True)
129
  return _indic_tts_model, _indic_tts_tokenizer, _indic_desc_tokenizer
@@ -145,11 +148,16 @@ def kokoro_tts_to_wav(text: str, speed: float = 1.0):
145
  model, tokenizer, desc_tokenizer = _get_indic_tts()
146
  device = model.device
147
 
148
- input_ids = desc_tokenizer(INDIC_TTS_DESCRIPTION, return_tensors="pt").input_ids.to(device)
149
- prompt_ids = tokenizer(text, return_tensors="pt").input_ids.to(device)
150
 
151
  with torch.no_grad():
152
- generation = model.generate(input_ids=input_ids, prompt_input_ids=prompt_ids)
 
 
 
 
 
153
 
154
  audio_np = generation.cpu().numpy().squeeze()
155
  sample_rate = model.config.sampling_rate
 
121
  _indic_tts_model = ParlerTTSForConditionalGeneration.from_pretrained(
122
  "ai4bharat/indic-parler-tts"
123
  ).to(device)
124
+ # Prompt tokenizer = the model repo's tokenizer (tokenizes the text to speak)
125
  _indic_tts_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
126
+ # Description tokenizer = the text-encoder's tokenizer (tokenizes the voice description).
127
+ # Loaded from text_encoder._name_or_path per the official AI4Bharat example — NOT a subfolder.
128
  _indic_desc_tokenizer = AutoTokenizer.from_pretrained(
129
+ _indic_tts_model.config.text_encoder._name_or_path
130
  )
131
  print(f"[IndicTTS] Model ready on {device}.", flush=True)
132
  return _indic_tts_model, _indic_tts_tokenizer, _indic_desc_tokenizer
 
148
  model, tokenizer, desc_tokenizer = _get_indic_tts()
149
  device = model.device
150
 
151
+ desc_inputs = desc_tokenizer(INDIC_TTS_DESCRIPTION, return_tensors="pt").to(device)
152
+ prompt_inputs = tokenizer(text, return_tensors="pt").to(device)
153
 
154
  with torch.no_grad():
155
+ generation = model.generate(
156
+ input_ids=desc_inputs.input_ids,
157
+ attention_mask=desc_inputs.attention_mask,
158
+ prompt_input_ids=prompt_inputs.input_ids,
159
+ prompt_attention_mask=prompt_inputs.attention_mask,
160
+ )
161
 
162
  audio_np = generation.cpu().numpy().squeeze()
163
  sample_rate = model.config.sampling_rate
requirements.txt CHANGED
@@ -1,7 +1,7 @@
1
  --extra-index-url https://download.pytorch.org/whl/cu121
2
  torch==2.3.1+cu121
3
  torchaudio==2.3.1+cu121
4
- git+https://github.com/huggingface/parler-tts.git
5
  transformers==4.46.1
6
  hf_transfer
7
  huggingface_hub>=0.23.2,<1.0
 
1
  --extra-index-url https://download.pytorch.org/whl/cu121
2
  torch==2.3.1+cu121
3
  torchaudio==2.3.1+cu121
4
+ git+https://github.com/huggingface/parler-tts.git@d108732cd57788ec86bc857d99a6cabd66663d68
5
  transformers==4.46.1
6
  hf_transfer
7
  huggingface_hub>=0.23.2,<1.0