Wav2Vec2-XLS-R-300M-21-EN
Facebook's Wav2Vec2 XLS-R fine-tuned for Speech Translation.
This is a SpeechEncoderDecoderModel model.
The encoder was warm-started from the facebook/wav2vec2-xls-r-300m checkpoint and
the decoder from the facebook/mbart-large-50 checkpoint.
Consequently, the encoder-decoder model was fine-tuned on 21 {lang} -> en translation pairs of the Covost2 dataset.
The model can translate from the following spoken languages {lang} -> en (English):
{fr, de, es, ca, it, ru, zh-CN, pt, fa, et, mn, nl, tr, ar, sv-SE, lv, sl, ta, ja, id, cy} -> en
For more information, please refer to Section 5.1.2 of the official XLS-R paper.
Usage
Demo
The model can be tested directly on the speech recognition widget on this model card! Simple record some audio in one of the possible spoken languages or pick an example audio file to see how well the checkpoint can translate the input.
Example
As this a standard sequence to sequence transformer model, you can use the generate method to generate the
transcripts by passing the speech features to the model.
You can use the model directly via the ASR pipeline
from datasets import load_dataset
from transformers import pipeline
# replace following lines to load an audio file of your choice
librispeech_en = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
audio_file = librispeech_en[0]["file"]
asr = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-xls-r-300m-21-to-en", feature_extractor="facebook/wav2vec2-xls-r-300m-21-to-en")
translation = asr(audio_file)
or step-by-step as follows:
import torch
from transformers import Speech2Text2Processor, SpeechEncoderDecoderModel
from datasets import load_dataset
model = SpeechEncoderDecoderModel.from_pretrained("facebook/wav2vec2-xls-r-300m-21-to-en")
processor = Speech2Text2Processor.from_pretrained("facebook/wav2vec2-xls-r-300m-21-to-en")
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
inputs = processor(ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["array"]["sampling_rate"], return_tensors="pt")
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
transcription = processor.batch_decode(generated_ids)
Results {lang} -> en
See the row of XLS-R (0.3B) for the performance on Covost2 for this model.
