Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import subprocess
|
2 |
+
subprocess.run(["pip", "install", "gradio", "--upgrade"])
|
3 |
+
subprocess.run(["pip", "install", "transformers"])
|
4 |
+
subprocess.run(["pip", "install", "torchaudio", "--upgrade"])
|
5 |
+
|
6 |
+
import gradio as gr
|
7 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
8 |
+
import torchaudio
|
9 |
+
import torch
|
10 |
+
|
11 |
+
# Load model and processor
|
12 |
+
processor = Wav2Vec2Processor.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-italian")
|
13 |
+
model = Wav2Vec2ForCTC.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-italian")
|
14 |
+
|
15 |
+
# Function to perform ASR on audio data
|
16 |
+
def transcribe_audio(audio_data):
|
17 |
+
print("Received audio data:", audio_data) # Debug print
|
18 |
+
|
19 |
+
# Check if audio_data is None or not a tuple of length 2
|
20 |
+
if audio_data is None or not isinstance(audio_data, tuple) or len(audio_data) != 2:
|
21 |
+
return "Invalid audio data format."
|
22 |
+
|
23 |
+
sample_rate, waveform = audio_data
|
24 |
+
|
25 |
+
# Check if waveform is None or not a NumPy array
|
26 |
+
if waveform is None or not isinstance(waveform, torch.Tensor):
|
27 |
+
return "Invalid audio data format."
|
28 |
+
|
29 |
+
try:
|
30 |
+
# Convert audio data to mono and normalize
|
31 |
+
audio_data = torchaudio.transforms.Resample(sample_rate, 100000)(waveform)
|
32 |
+
audio_data = torchaudio.functional.gain(audio_data, gain_db=5.0)
|
33 |
+
|
34 |
+
# Apply custom preprocessing to the audio data if needed
|
35 |
+
input_values = processor(audio_data[0], return_tensors="pt").input_values
|
36 |
+
|
37 |
+
# Perform ASR
|
38 |
+
with torch.no_grad():
|
39 |
+
logits = model(input_values).logits
|
40 |
+
|
41 |
+
# Decode the output
|
42 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
43 |
+
transcription = processor.batch_decode(predicted_ids)
|
44 |
+
|
45 |
+
return transcription[0]
|
46 |
+
|
47 |
+
except Exception as e:
|
48 |
+
return f"An error occurred: {str(e)}"
|
49 |
+
|
50 |
+
# Create Gradio interface
|
51 |
+
audio_input = gr.Audio(sources=["microphone"])
|
52 |
+
gr.Interface(fn=transcribe_audio, inputs=audio_input, outputs="text").launch()
|