File size: 2,288 Bytes
dfbe659 b6fd1ba dfbe659 c6f025d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 |
---
language: it
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in it (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-it")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-it")
# load dataset
ds = load_dataset("common_voice", "it", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|