S2T-SMALL-MUSTC-EN-ES-ST

s2t-small-mustc-en-es-st is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST). The S2T model was proposed in this paper and released in this repository

Model description

S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the transcripts/translations autoregressively.

Intended uses & limitations

This model can be used for end-to-end English speech to Spanish text translation. See the model hub to look for other S2T checkpoints.

How to use

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.

Note: The Speech2TextProcessor object uses torchaudio to extract the filter bank features. Make sure to install the torchaudio package before running this example.

You could either install those as extra speech dependancies with pip install transformers"[speech, sentencepiece]" or install the packages seperatly with pip install torchaudio sentencepiece.

import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf

model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-es-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-es-st")

def map_to_array(batch):
    speech, _ = sf.read(batch["file"])
    batch["speech"] = speech
    return batch

ds = load_dataset(
    "patrickvonplaten/librispeech_asr_dummy",
    "clean",
    split="validation"
)
ds = ds.map(map_to_array)

inputs = processor(
    ds["speech"][0],
    sampling_rate=16_000,
    return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])

translation = processor.batch_decode(generated_ids, skip_special_tokens=True)

Training data

The s2t-small-mustc-en-es-st is trained on English-Spanish subset of MuST-C. MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems for speech translation from English into several languages. For each target language, MuST-C comprises several hundred hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual transcriptions and translations.

Training procedure

Preprocessing

The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization) is applied to each example.

The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.

Training

The model is trained with standard autoregressive cross-entropy loss and using SpecAugment. The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate model training and for better performance the encoder is pre-trained for English ASR.

Evaluation results

MuST-C test results for en-es (BLEU score): 27.2

BibTeX entry and citation info

@inproceedings{wang2020fairseqs2t,
  title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
  author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
  booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
  year = {2020},
}