Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -6,28 +6,32 @@ from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
|
|
6 |
import torch
|
7 |
import torchaudio
|
8 |
|
|
|
|
|
|
|
9 |
app = FastAPI()
|
10 |
|
|
|
11 |
processor = Wav2Vec2Processor.from_pretrained("tacab/ASR_SOMALI")
|
12 |
model = Wav2Vec2ForCTC.from_pretrained("tacab/ASR_SOMALI")
|
13 |
model.to("cpu")
|
14 |
|
15 |
@app.post("/transcribe")
|
16 |
async def transcribe(file: UploadFile = File(...)):
|
|
|
17 |
audio_bytes = await file.read()
|
18 |
-
|
19 |
temp_path = "/tmp/temp.wav"
|
20 |
with open(temp_path, "wb") as f:
|
21 |
f.write(audio_bytes)
|
22 |
|
|
|
23 |
speech_array, sampling_rate = torchaudio.load(temp_path)
|
24 |
-
inputs = processor(speech_array.squeeze(), return_tensors="pt", sampling_rate=sampling_rate)
|
25 |
|
|
|
|
|
26 |
with torch.no_grad():
|
27 |
logits = model(**inputs).logits
|
28 |
-
|
29 |
predicted_ids = torch.argmax(logits, dim=-1)
|
30 |
transcription = processor.batch_decode(predicted_ids)[0]
|
31 |
|
32 |
return {"text": transcription}
|
33 |
-
|
|
|
6 |
import torch
|
7 |
import torchaudio
|
8 |
|
9 |
+
# β
Explicitly set the backend so .wav files load properly
|
10 |
+
torchaudio.set_audio_backend("soundfile")
|
11 |
+
|
12 |
app = FastAPI()
|
13 |
|
14 |
+
# β
Load model and processor
|
15 |
processor = Wav2Vec2Processor.from_pretrained("tacab/ASR_SOMALI")
|
16 |
model = Wav2Vec2ForCTC.from_pretrained("tacab/ASR_SOMALI")
|
17 |
model.to("cpu")
|
18 |
|
19 |
@app.post("/transcribe")
|
20 |
async def transcribe(file: UploadFile = File(...)):
|
21 |
+
# β
Save uploaded file to /tmp
|
22 |
audio_bytes = await file.read()
|
|
|
23 |
temp_path = "/tmp/temp.wav"
|
24 |
with open(temp_path, "wb") as f:
|
25 |
f.write(audio_bytes)
|
26 |
|
27 |
+
# β
Load audio file
|
28 |
speech_array, sampling_rate = torchaudio.load(temp_path)
|
|
|
29 |
|
30 |
+
# β
Run through ASR model
|
31 |
+
inputs = processor(speech_array.squeeze(), return_tensors="pt", sampling_rate=sampling_rate)
|
32 |
with torch.no_grad():
|
33 |
logits = model(**inputs).logits
|
|
|
34 |
predicted_ids = torch.argmax(logits, dim=-1)
|
35 |
transcription = processor.batch_decode(predicted_ids)[0]
|
36 |
|
37 |
return {"text": transcription}
|
|