update readme
Browse files
README.md
CHANGED
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
- de
|
5 |
+
- nl
|
6 |
+
- es
|
7 |
+
- fr
|
8 |
+
- it
|
9 |
+
- pt
|
10 |
+
- ro
|
11 |
+
- ru
|
12 |
+
datasets:
|
13 |
+
- mustc
|
14 |
+
tags:
|
15 |
+
- audio
|
16 |
+
- speech-translation
|
17 |
+
- automatic-speech-recognition
|
18 |
+
license: MIT
|
19 |
+
---
|
20 |
+
|
21 |
+
|
22 |
+
# S2T-MEDIUM-MUSTC-MULTILINGUAL-ST
|
23 |
+
|
24 |
+
`s2t-medium-mustc-multilingual-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Multilingual Speech Translation (ST).
|
25 |
+
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
|
26 |
+
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
|
27 |
+
|
28 |
+
|
29 |
+
## Model description
|
30 |
+
|
31 |
+
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
|
32 |
+
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
|
33 |
+
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
|
34 |
+
transcripts/translations autoregressively.
|
35 |
+
|
36 |
+
## Intended uses & limitations
|
37 |
+
|
38 |
+
This model can be used for end-to-end English speech to French text translation.
|
39 |
+
See the [model hub](https://huggingface.co/models?filter=speech_to_text_transformer) to look for other S2T checkpoints.
|
40 |
+
|
41 |
+
|
42 |
+
### How to use
|
43 |
+
|
44 |
+
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
|
45 |
+
transcripts by passing the speech features to the model.
|
46 |
+
|
47 |
+
For multilingual speech translation models, `eos_token_id` is used as the `decoder_start_token_id` and
|
48 |
+
the target language id is forced as the first generated token. To force the target language id as the first
|
49 |
+
generated token, pass the `forced_bos_token_id` parameter to the `generate()` method. The following
|
50 |
+
example shows how to transate English speech to French and German text using the `facebook/s2t-medium-mustc-multilingual-st`
|
51 |
+
checkpoint.
|
52 |
+
|
53 |
+
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
|
54 |
+
filter bank features. Make sure to install the `torchaudio` package before running this example.*
|
55 |
+
|
56 |
+
You could either install those as extra speech dependancies with
|
57 |
+
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
|
58 |
+
with `pip install torchaudio sentencepiece`.
|
59 |
+
|
60 |
+
|
61 |
+
```python
|
62 |
+
import torch
|
63 |
+
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
|
64 |
+
from datasets import load_dataset
|
65 |
+
import soundfile as sf
|
66 |
+
|
67 |
+
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-medium-mustc-multilingual-st")
|
68 |
+
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-medium-mustc-multilingual-st")
|
69 |
+
|
70 |
+
def map_to_array(batch):
|
71 |
+
speech, _ = sf.read(batch["file"])
|
72 |
+
batch["speech"] = speech
|
73 |
+
return batch
|
74 |
+
|
75 |
+
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
|
76 |
+
ds = ds.map(map_to_array)
|
77 |
+
|
78 |
+
inputs = processor(ds["speech"][0], sampling_rate=16_000, return_tensors="pt")
|
79 |
+
|
80 |
+
# translate English Speech To French Text
|
81 |
+
generated_ids = model.generate(
|
82 |
+
input_ids=inputs["input_features"],
|
83 |
+
attention_mask=inputs["attention_mask"],
|
84 |
+
forced_bos_token_id=processor.tokenizer.lang_code_to_id["fr"]
|
85 |
+
)
|
86 |
+
translation_fr = processor.batch_decode(generated_ids)
|
87 |
+
|
88 |
+
# translate English Speech To German Text
|
89 |
+
generated_ids = model.generate(
|
90 |
+
input_ids=inputs["input_features"],
|
91 |
+
attention_mask=inputs["attention_mask"],
|
92 |
+
forced_bos_token_id=processor.tokenizer.lang_code_to_id["de"]
|
93 |
+
)
|
94 |
+
translation_de = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
95 |
+
```
|
96 |
+
|
97 |
+
|
98 |
+
## Training data
|
99 |
+
|
100 |
+
The s2t-medium-mustc-multilingual-st is trained on [MuST-C](https://ict.fbk.eu/must-c/).
|
101 |
+
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
|
102 |
+
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
|
103 |
+
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
|
104 |
+
transcriptions and translations.
|
105 |
+
|
106 |
+
|
107 |
+
## Training procedure
|
108 |
+
|
109 |
+
### Preprocessing
|
110 |
+
|
111 |
+
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
|
112 |
+
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
|
113 |
+
is applied to each example.
|
114 |
+
|
115 |
+
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 10,000.
|
116 |
+
|
117 |
+
|
118 |
+
### Training
|
119 |
+
|
120 |
+
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
|
121 |
+
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
|
122 |
+
model training and for better performance the encoder is pre-trained for multilingual ASR. For multilingual models, target language ID token
|
123 |
+
is used as target BOS.
|
124 |
+
|
125 |
+
## Evaluation results
|
126 |
+
|
127 |
+
MuST-C test results (BLEU score):
|
128 |
+
|
129 |
+
| En-De | En-Nl | En-Es | En-Fr | En-It | En-Pt | En-Ro | En-Ru |
|
130 |
+
|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|
|
131 |
+
| 24.5 | 28.6 | 28.2 | 34.9 | 24.6 | 31.1 | 23.8 | 16.0 |
|
132 |
+
|
133 |
+
|
134 |
+
|
135 |
+
### BibTeX entry and citation info
|
136 |
+
|
137 |
+
```bibtex
|
138 |
+
@inproceedings{wang2020fairseqs2t,
|
139 |
+
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
|
140 |
+
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
|
141 |
+
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
|
142 |
+
year = {2020},
|
143 |
+
}
|
144 |
+
|
145 |
+
```
|