Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torchaudio
|
| 3 |
+
from transformers import pipeline
|
| 4 |
+
|
| 5 |
+
# Load only the Moul-Sout-100 model
|
| 6 |
+
asr_pipeline = pipeline("automatic-speech-recognition", model="01Yassine/moul-sout-100")
|
| 7 |
+
|
| 8 |
+
# Adjust generation config if necessary
|
| 9 |
+
asr_pipeline.model.generation_config.input_ids = asr_pipeline.model.generation_config.forced_decoder_ids
|
| 10 |
+
asr_pipeline.model.generation_config.forced_decoder_ids = None
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
def ensure_mono_16k(audio_path):
|
| 14 |
+
"""Load audio, convert to mono + 16kHz, and save a temp version."""
|
| 15 |
+
waveform, sr = torchaudio.load(audio_path)
|
| 16 |
+
|
| 17 |
+
# Convert to mono if necessary
|
| 18 |
+
if waveform.shape[0] > 1:
|
| 19 |
+
waveform = waveform.mean(dim=0, keepdim=True)
|
| 20 |
+
|
| 21 |
+
# Resample to 16kHz if necessary
|
| 22 |
+
if sr != 16000:
|
| 23 |
+
resampler = torchaudio.transforms.Resample(sr, 16000)
|
| 24 |
+
waveform = resampler(waveform)
|
| 25 |
+
sr = 16000
|
| 26 |
+
|
| 27 |
+
tmp_path = "/tmp/processed_16k.wav"
|
| 28 |
+
torchaudio.save(tmp_path, waveform, sr)
|
| 29 |
+
return tmp_path
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def transcribe(audio):
|
| 33 |
+
if audio is None:
|
| 34 |
+
return "Please record or upload an audio file."
|
| 35 |
+
|
| 36 |
+
# Process and transcribe
|
| 37 |
+
processed_audio = ensure_mono_16k(audio)
|
| 38 |
+
result = asr_pipeline(processed_audio)["text"]
|
| 39 |
+
|
| 40 |
+
return result
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
title = "ποΈ Moul-Sout ASR π²π¦"
|
| 44 |
+
description = """
|
| 45 |
+
**Moul-Sout** model for Darija ASR π²π¦.
|
| 46 |
+
You can record or upload an audio sample (it will be automatically resampled to 16 kHz mono),
|
| 47 |
+
and view the transcription result below.
|
| 48 |
+
"""
|
| 49 |
+
|
| 50 |
+
with gr.Blocks(title=title) as demo:
|
| 51 |
+
gr.Markdown(f"# {title}\n{description}")
|
| 52 |
+
|
| 53 |
+
with gr.Row():
|
| 54 |
+
audio_input = gr.Audio(
|
| 55 |
+
sources=["microphone", "upload"],
|
| 56 |
+
type="filepath",
|
| 57 |
+
label="π€ Record or Upload Audio (auto 16 kHz mono)"
|
| 58 |
+
)
|
| 59 |
+
|
| 60 |
+
transcribe_btn = gr.Button("π Transcribe")
|
| 61 |
+
|
| 62 |
+
output_text = gr.Textbox(label="π© Transcription Output")
|
| 63 |
+
|
| 64 |
+
transcribe_btn.click(
|
| 65 |
+
fn=transcribe,
|
| 66 |
+
inputs=[audio_input],
|
| 67 |
+
outputs=[output_text]
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
# Local launch
|
| 71 |
+
if __name__ == "__main__":
|
| 72 |
+
demo.launch()
|