ovieyra21 commited on
Commit
511d4bd
β€’
1 Parent(s): a845bee

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +48 -78
app.py CHANGED
@@ -1,101 +1,71 @@
 
 
1
  import torch
 
 
2
  from transformers import pipeline
3
- from transformers.pipelines.audio_utils import ffmpeg_read
4
- import gradio as gr
5
 
6
- MODEL_NAME = "openai/whisper-small"
7
- BATCH_SIZE = 8
8
 
9
- device = 0 if torch.cuda.is_available() else "cpu"
 
 
 
 
 
 
 
 
10
 
11
- pipe = pipeline(
12
- task="automatic-speech-recognition",
13
- model=MODEL_NAME,
14
- chunk_length_s=30,
15
- device=device,
16
- )
17
 
18
 
19
- # Copied from https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/utils.py#L50
20
- def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."):
21
- if seconds is not None:
22
- milliseconds = round(seconds * 1000.0)
23
 
24
- hours = milliseconds // 3_600_000
25
- milliseconds -= hours * 3_600_000
 
26
 
27
- minutes = milliseconds // 60_000
28
- milliseconds -= minutes * 60_000
29
 
30
- seconds = milliseconds // 1_000
31
- milliseconds -= seconds * 1_000
 
 
 
32
 
33
- hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
34
- return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
35
- else:
36
- # we have a malformed timestamp so just return it as is
37
- return seconds
38
 
 
 
 
 
 
39
 
40
- def transcribe(file, task, return_timestamps):
41
- outputs = pipe(file, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=return_timestamps)
42
- text = outputs["text"]
43
- if return_timestamps:
44
- timestamps = outputs["chunks"]
45
- timestamps = [
46
- f"[{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}"
47
- for chunk in timestamps
48
- ]
49
- text = "\n".join(str(feature) for feature in timestamps)
50
- return text
51
 
 
 
 
 
 
 
52
 
53
  demo = gr.Blocks()
54
 
55
- mic_transcribe = gr.Interface(
56
- fn=transcribe,
57
- inputs=[
58
- gr.inputs.Audio(source="microphone", type="filepath", optional=True),
59
- gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
60
- gr.inputs.Checkbox(default=False, label="Return timestamps"),
61
- ],
62
- outputs="text",
63
- layout="horizontal",
64
- theme="huggingface",
65
- title="Whisper Demo: Transcribe Audio",
66
- description=(
67
- "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the"
68
- f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files"
69
- " of arbitrary length."
70
- ),
71
- allow_flagging="never",
72
  )
73
 
74
- file_transcribe = gr.Interface(
75
- fn=transcribe,
76
- inputs=[
77
- gr.inputs.Audio(source="upload", optional=True, label="Audio file", type="filepath"),
78
- gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
79
- gr.inputs.Checkbox(default=False, label="Return timestamps"),
80
- ],
81
- outputs="text",
82
- layout="horizontal",
83
- theme="huggingface",
84
- title="Whisper Demo: Transcribe Audio",
85
- description=(
86
- "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the"
87
- f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files"
88
- " of arbitrary length."
89
- ),
90
- examples=[
91
- ["./example.flac", "transcribe", False],
92
- ["./example.flac", "transcribe", True],
93
- ],
94
- cache_examples=True,
95
- allow_flagging="never",
96
  )
97
 
98
  with demo:
99
- gr.TabbedInterface([mic_transcribe, file_transcribe], ["Transcribe Microphone", "Transcribe Audio File"])
100
 
101
- demo.launch(enable_queue=True)
 
1
+ import gradio as gr
2
+ import numpy as np
3
  import torch
4
+ from datasets import load_dataset
5
+
6
  from transformers import pipeline
7
+ from transformers import VitsModel, VitsTokenizer
 
8
 
 
 
9
 
10
+ device = "cuda:0" if torch.cuda.is_available() else "cpu"
11
+
12
+ # load speech translation checkpoint
13
+ asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
14
+
15
+
16
+
17
+ model = VitsModel.from_pretrained("facebook/mms-tts-spa")
18
+ processor = VitsTokenizer.from_pretrained("facebook/mms-tts-spa")
19
 
 
 
 
 
 
 
20
 
21
 
 
 
 
 
22
 
23
+ def translate(audio):
24
+ outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"language": "es","task": "transcribe"})
25
+ return outputs["text"]
26
 
 
 
27
 
28
+ def synthesise(text):
29
+ inputs = processor(text=text, return_tensors="pt")
30
+ with torch.no_grad():
31
+ speech = model(inputs["input_ids"].to(device))
32
+ return speech.audio[0]
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
+ ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
47
+ """
48
 
49
  demo = gr.Blocks()
50
 
51
+ mic_translate = gr.Interface(
52
+ fn=speech_to_speech_translation,
53
+ inputs=gr.Audio(source="microphone", type="filepath"),
54
+ outputs=gr.Audio(label="Generated Speech", type="numpy"),
55
+ title=title,
56
+ description=description,
 
 
 
 
 
 
 
 
 
 
 
57
  )
58
 
59
+ file_translate = gr.Interface(
60
+ fn=speech_to_speech_translation,
61
+ inputs=gr.Audio(source="upload", type="filepath"),
62
+ outputs=gr.Audio(label="Generated Speech", type="numpy"),
63
+ examples=[["./example.wav"]],
64
+ title=title,
65
+ description=description,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
  )
67
 
68
  with demo:
69
+ gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
70
 
71
+ demo.launch()