shniranjan's picture
Update README.md
98c6744
|
raw
history blame
974 Bytes
## Usage
The model can be used directly (without a language model) as follows:
---
language:
- ne
tags:
- speech-to-text
---
```python
import soundfile as sf
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import argparse
def parse_transcription(wav_file):
# load pretrained model
processor = Wav2Vec2Processor.from_pretrained("shniranjan/wav2vec2-large-xlsr-300m-nepali")
model = Wav2Vec2ForCTC.from_pretrained("shniranjan/wav2vec2-large-xlsr-300m-nepali")
# load audio
audio_input, sample_rate = sf.read(wav_file)
# pad input values and return pt tensor
input_values = processor(audio_input, sampling_rate=sample_rate, return_tensors="pt").input_values
# INFERENCE
# retrieve logits & take argmax
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
# transcribe
transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
print(transcription)
```