wilton commited on
Commit
95149f6
1 Parent(s): 340e344

updating of functions, new MMS model for spanish TTS support

Browse files
Files changed (2) hide show
  1. app.py +19 -19
  2. requirements.txt +1 -1
app.py CHANGED
@@ -2,47 +2,47 @@ import gradio as gr
2
  import numpy as np
3
  import torch
4
  from datasets import load_dataset
5
-
6
- from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
7
-
8
 
9
  device = "cuda:0" if torch.cuda.is_available() else "cpu"
10
 
11
  # load speech translation checkpoint
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")
21
- speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
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
 
29
  def synthesise(text):
30
- inputs = processor(text=text, return_tensors="pt")
31
- speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
32
- return speech.cpu()
 
33
 
34
 
35
  def speech_to_speech_translation(audio):
36
  translated_text = translate(audio)
37
  synthesised_speech = synthesise(translated_text)
38
- synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
39
- return 16000, synthesised_speech
40
 
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")
48
  """
@@ -69,4 +69,4 @@ file_translate = gr.Interface(
69
  with demo:
70
  gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
71
 
72
- demo.launch()
 
2
  import numpy as np
3
  import torch
4
  from datasets import load_dataset
5
+ from transformers import pipeline, VitsModel, AutoTokenizer
 
 
6
 
7
  device = "cuda:0" if torch.cuda.is_available() else "cpu"
8
 
9
  # load speech translation checkpoint
10
+ asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-small", device=device)
11
 
12
+ # load facebook mms espanish model/checkpoint
13
+ model = VitsModel.from_pretrained("facebook/mms-tts-spa")
14
+ tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-spa")
15
 
 
 
16
 
17
+ target_dtype = np.int16
18
+ max_range = np.iinfo(target_dtype).max
19
 
20
 
21
  def translate(audio):
22
+ outputs = asr_pipe(
23
+ audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "es"}
24
+ )
25
  return outputs["text"]
26
 
27
 
28
  def synthesise(text):
29
+ inputs = tokenizer(text, return_tensors="pt")
30
+ with torch.no_grad():
31
+ speech = model(**inputs).waveform
32
+ return speech.squeeze(0).cpu()
33
 
34
 
35
  def speech_to_speech_translation(audio):
36
  translated_text = translate(audio)
37
  synthesised_speech = synthesise(translated_text)
38
+ synthesised_speech = (synthesised_speech.numpy() * max_range).astype(np.int16)
39
+ return 16_000, synthesised_speech
40
 
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 *Spanish*. Demo uses OpenAI's [Whisper Small](https://huggingface.co/openai/whisper-small) model for speech translation, and Meta's
45
+ [MMS TTS Spanish](https://huggingface.co/facebook/mms-tts-spa) 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")
48
  """
 
69
  with demo:
70
  gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
71
 
72
+ demo.launch(debug=True)
requirements.txt CHANGED
@@ -1,4 +1,4 @@
1
  torch
2
- git+https://github.com/huggingface/transformers
3
  datasets
4
  sentencepiece
 
1
  torch
2
+ transformers
3
  datasets
4
  sentencepiece