Sagicc commited on
Commit
ca8742a
1 Parent(s): 4c01452

Update app.py

Browse files

Update TTS Model for Audio ML Course demo

Files changed (1) hide show
  1. app.py +4 -4
app.py CHANGED
@@ -12,9 +12,9 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
12
  asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
13
 
14
  # load text-to-speech checkpoint and speaker embeddings
15
- processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
16
 
17
- model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
18
  vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
19
 
20
  embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
@@ -22,7 +22,7 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
22
 
23
 
24
  def translate(audio):
25
- outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
26
  return outputs["text"]
27
 
28
 
@@ -41,7 +41,7 @@ def speech_to_speech_translation(audio):
41
 
42
  title = "Cascaded STST"
43
  description = """
44
- Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
45
  [SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
46
 
47
  ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 
12
  asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
13
 
14
  # load text-to-speech checkpoint and speaker embeddings
15
+ processor = SpeechT5Processor.from_pretrained("Sagicc/speecht5_finetuned_multilingual_librispeech_pl")
16
 
17
+ model = SpeechT5ForTextToSpeech.from_pretrained("Sagicc/speecht5_finetuned_multilingual_librispeech_pl").to(device)
18
  vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
19
 
20
  embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 
22
 
23
 
24
  def translate(audio):
25
+ outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "pl"})
26
  return outputs["text"]
27
 
28
 
 
41
 
42
  title = "Cascaded STST"
43
  description = """
44
+ Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Polish. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
45
  [SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
46
 
47
  ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")