A PyTorch implementation of paper "Learning Shared Semantic Space for Speech-to-Text Translation", ACL (Findings) 2021

Overview

Chimera: Learning Shared Semantic Space for Speech-to-Text Translation


This is a Pytorch implementation for the "Chimera" paper Learning Shared Semantic Space for Speech-to-Text Translation https://arxiv.org/abs/2105.03095 (accepted by ACL Findings 2021), which aims to bridge the modality gap by unifying the task of MT (textual Machine Translation) and ST (Speech-to-Text Translation). It has achieved new SOTA performance on all 8 language pairs in MuST-C benchmark, by utilizing an external MT corpus.


This repository is up to now a nightly version, and is bug-prone because of code refactoring. Also it is not fully tested on configurations other than the authors' working environment yet. However, we encourage you to first have a look at the results and model codes to get a general impression of what this project is about.

The code base is forked from FairSeq repository https://github.com/pytorch/fairseq.git (without an actual forking operation) in Septempber 2020. It than lags behind the later updates in FairSeq, and both the codes and checkpoints are not compatible with currect Fairseq version. You will need to modify the model codes for checkpoint configurations if you want to follow the new FairSeq codes.

CONTRIBUTION: You are also more than welcomed to test our code on your machines, and report feedbacks on results, bugs and performance!



Results

Our model (Chimera) achieves new state-of-the-art results on all 8 language pairs on MuST-C:

Direction EN-DE EN-FR EN-RU EN-ES EN-IT EN-RO EN-PT EN-NL
BLEU 26.3 35.6 17.4 30.6 25.0 24.0 30.2 29.2

Chimera novelly learns M distinct "memories" to store specific types of semantic information from both audio and text inputs. Shown below is a visualization of the "Memories" learned by Chimera-16, which is a variant with M = 16. Each learned cluster represents a individual type of information, while each marker is a sentence sample. "+" and "." means text and audio samples, respectively.

We can see more clearly from below (left) that memories learn a well-clustered semantic space, forming a "semantic" alignment (rather than spatial) between audio and text inputs, while ignoring the modality differences.

On the right, we zoom in to focus one cluster in specific, and it can be easily observed that the vectors are well structured as well, with inputs with (probably one of) similar semantic features close in space to each other.

We can even focus on one instance of translation, and see how the memories works. Shown below visualizes the alignment between audio attention and text attention, which tightly gather around the diagonal line. Different colors represents different memories, which attend to different semantic segments of sentence / audio as shown in the figure.



Trained Checkpoints

Our trained checkpoints are available at:

Translation Direction filename External url
English-to-Deutsch Chimera_EN2DE.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2DE.pt
English-to-French Chimera_EN2FR.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2FR.pt
English-to-Russian Chimera_EN2RU.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2RU.pt
English-to-Espanol Chimera_EN2ES.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2ES.pt
English-to-Italiano Chimera_EN2IT.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2IT.pt
English-to-Romanian Chimera_EN2RO.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2RO.pt
English-to-Portuguese Chimera_EN2PT.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2PT.pt
English-to-Dutch Chimera_EN2NL.pt http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/acl2021/chimera/Chimera_EN2NL.pt



Interactive Translation

You can download any one checkpoint mentioned above to local, and translate local audios (only .wav files supported) to another language! To do this, you only need to run the model in an interactive mode. For example, you want to translate from English to Deutsh (DE) with an already trained checkpoint at $CHECKPOINT:

bash run.sh --script chimera/scripts/interactive-en2any-ST.sh \
    --target de --checkpoint $CHECKPOINT

The program will prompt an input file name like this:

2021-04-02 10:00:00 | INFO | fairseq_cli.interactive | Type the input sentence and press return:

After inputing the file name, the program will translate outputs like:

H-0     -1.0      ▁Nach ▁dem ...
D-0     -1.0      Nach dem ...
P-0     -1.0000 -1.0000 ...

NOTE: Do not input a file too large. Normally the model can translate 1~5 normal-length sentences in one time. If the input sentence is too long, the program could crash.

To exit the interactive mode, you only need to input an invalid file name.

To translate to other languages, remember to replace de with their language codes (in lower case):

Language Code
Deutsch (German) DE / de
French FR / fr
Espanol (Spanish) ES / es
Russian RU / ru
Italiano (Italian) IT / it
Romanian RO / ro
Portuguese PT / pt
Dutch (Netherlands) NL / nl



Training a Model on MuST-C

Let's first take a look at training an English-to-Deutsch model as an example.

Data Preparation

  1. Prerequisites and Configuration First check that requirements are met for pip in requirements.txt and for apt in apt-requirements.txt. Some items in the two files may be redundant, but we haven't got time to check and eliminate them.

For configuration, please set the global variables of $WMT_ROOT, $MUSTC_ROOT and SAVE_ROOT These will be where to put the datasets and checkpoints. For example:

export MUSTC_ROOT="speech_data/mustc"
export WMT_ROOT="wmt_data"
export SAVE_ROOT="checkpoints"
export target=de
mkdir -p $MUSTC_ROOT $WMT_ROOT $SAVE_ROOT

NOTE: This simple configuration is a prerequisite for most of the following steps. Here export target=de means the translation direction is English to Deutsch.

  1. Download and uncompress the EN-to-DE MuST-C dataset to $MUSTC_ROOT/en-$target. TIP: to speed up uncompressing a file too large, you can replace tar xzvf with: pigz -dc $TARFILE | tar xvf -

  2. Download the WMT to $WMT_ROOT/orig via:

bash chimera/prepare_data/download-wmt.sh --wmt14 --data-dir $WMT_ROOT --target $target

This may sometimes be too slow as the connection to statmt.org is not steady in some places. In this case you can turn to other faster download sources if possible.

  1. Append MuST-C text data to $WMT_ROOT, and prepare the datasets and produce a joint spm dictionary:
bash chimera/prepare_data/prepare-wmt-en2any.sh \
    --data-dir $WMT_ROOT --wmt14 --original-dev \
    --external mustc --target $target --subword spm
python3 chimera/prepare_data/prep_mustc_data.py \
    --data-root $MUSTC_ROOT --task wave \
    --ignore_fbank80 --joint_spm wmt14-en-$target-spm \
    --languages $target --vocab-type unigram --vocab-size 10000

NOTE: if the first command is executed correctly, you will see one line in the output:

Existing spm dictionary chimera/resources/wmt14-en-de-spm detected. Copying...

If not, the program will still produce one dictionary on the run and reports No existing spm detected. Learning unigram spm on wmt14_en_de/tmp/train.de-en ... This is okay in most cases, with the only risk being a potential mismatch to already trained checkpoints we provided.

Training

To reproduce the results in the last row in Figure 1 in paper, you can directly use the training scripts available as follows.

  1. Pre-training on MT data:
bash run.sh --script chimera/scripts/train-en2any-MT.sh \
    --target $target --dataset wmt14 --max_updates 500000

If you like, you can specify some arguments other than default values. The default setting is --seed 1 --num-gpus 8, which makes the command look like bash run.sh --script chimera/scripts/train-en2$target-MT.sh --seed 1 --num-gpus 8. Value for --num-gpus is recommended to be power of 2, and smaller than 8, e.g. {1, 2, 4, 8}.

  1. Fine-tuning on MuST-C data:
bash run.sh --script chimera/scripts/train-en2any-ST.sh \
    --target $target --dataset wmt14 --max_updates 150000

This script moves the MT-pre-trained model from ${MT_SAVE_DIR}/checkpoint_best.pt to ${ST_SAVE_DIR} as a initialization for ST fine-tuning.

Optionally, if you need to resume a single ST training, you can add argument --resume to the command to avoid overwriting the existing ${ST_SAVE_DIR}/checkpoint_last.pt.

The scripts in step 4 and 5 forks a separate background evaluation process while running. The process monitors $MT_SAVE_ROOT or $ST_SAVE_ROOT and evaluates any new checkpoints. Don't worry, it will be automatically killed after the training finishes, unless the script is Ctrl-C'ed, in which case, you can manually raise the suicide flag by touch chimera/tools/auto-generate-suicide.code to kill the background generation process.

Note that this automatic process only evaluates a single checkpoint (with no averaging), and with a low beam width.

  1. Averaging Checkpoints and Evaluate It

Suppose the best ST checkpoint is at epoch $BEST_EPOCH, and we want to averaging 7 checkpoints around it.

python3 chimera/tools/eval-average-checkpoint.py \
    --ckpt-dir $ST_SAVE_ROOT --number-of-ckpts 7 \
    --center-of-ckpts $BEST_EPOCH

Other Language Pairs

For language pairs English-to-{French, Russian, Espanol}, you only need to replace the export target=de with {fr, ru, es} in step 0, and then run the steps 1~5.

For language pairs English-to-{Italiano, Portuguese, Dutch, Romanian}, the MT data is different, so we need to modify Step 2 and 3. All other Steps remains unchanged.

English to Romanian

For Romanian, we use WMT16 corpora in our paper.

The Step 2 changes to

bash chimera/prepare_data/download-wmt.sh --wmt16 --data-dir $WMT_ROOT --target ro

Step 3 remains unchanged.

English to {Italiano, Portuguese, Dutch}

These language pairs uses OPUS100 as external MT corpora.

The Step 2 changes to

bash chimera/prepare_data/download-opus100.sh --data-dir $WMT_ROOT

Step 3 changes to

bash chimera/prepare_data/prepare-opus100-en2any.sh \
    --data-dir $WMT_ROOT --original-dev \
    --external mustc --target $target --subword spm
python3 chimera/prepare_data/prep_mustc_data.py \
    --data-root $MUSTC_ROOT --task wave \
    --ignore_fbank80 --joint_spm wmt14-en-$target-spm \
    --languages $target --vocab-type unigram --vocab-size 10000

Actually, only the first command of Step 3 changes.

Evaluating a Checkpoint

You can also manually evaluate the performance of any one checkpoint on MuST-C test set. Suppose the path to your checkpoint is $CHECKPOINT

target=de bash chimera/generate/generate-mustc-final.sh $CHECKPOINT



License

Part of codes (especially codes outside chimera/) is adapted from FAIRSEQ code base, therefore carrying the MIT License of its original codes. See NOTICE.md for more details.

Citation

Please cite as:

@article{han2021learning,
  title={Learning Shared Semantic Space for Speech-to-Text Translation},
  author={Han, Chi and Wang, Mingxuan and Ji, Heng and Li, Lei},
  journal={arXiv preprint arXiv:2105.03095},
  year={2021}
}
Owner
Chi Han
CS Graduate student at UIUC.
Chi Han
Extracting Summary Knowledge Graphs from Long Documents

GraphSum This repo contains the data and code for the G2G model in the paper: Extracting Summary Knowledge Graphs from Long Documents. The other basel

Zeqiu (Ellen) Wu 10 Oct 21, 2022
Coreference resolution for English, German and Polish, optimised for limited training data and easily extensible for further languages

Coreferee Author: Richard Paul Hudson, msg systems ag 1. Introduction 1.1 The basic idea 1.2 Getting started 1.2.1 English 1.2.2 German 1.2.3 Polish 1

msg systems ag 169 Dec 21, 2022
2021语言与智能技术竞赛:机器阅读理解任务

LICS2021 MRC 1. 项目&任务介绍 本项目基于官方给定的baseline(DuReader-Checklist-BASELINE)进行二次改造,对整个代码框架做了简单的重构,对核心网络结构添加了注释,解耦了数据读取的模块,并添加了阈值确认的功能,一些小的细节也做了改进。 本次任务为202

roar 29 Dec 05, 2022
List of GSoC organisations with number of times they have been selected.

Welcome to GSoC Organisation Frequency And Details 👋 List of GSoC organisations with number of times they have been selected, techonologies, topics,

Shivam Kumar Jha 41 Oct 01, 2022
Using Bert as the backbone model for lime, designed for NLP task explanation (sentence pair text classification task)

Lime Comparing deep contextualized model for sentences highlighting task. In addition, take the classic explanation model "LIME" with bert-base model

JHJu 2 Jan 18, 2022
Python functions for summarizing and improving voice dictation input.

Helpmespeak Help me speak uses Python functions for summarizing and improving voice dictation input. Get started with OpenAI gpt-3 OpenAI is a amazing

Margarita Humanitarian Foundation 6 Dec 17, 2022
Pytorch NLP library based on FastAI

Quick NLP Quick NLP is a deep learning nlp library inspired by the fast.ai library It follows the same api as fastai and extends it allowing for quick

Agis pof 283 Nov 21, 2022
Repository for Project Insight: NLP as a Service

Project Insight NLP as a Service Contents Introduction Features Installation Setup and Documentation Project Details Demonstration Directory Details H

Abhishek Kumar Mishra 286 Dec 06, 2022
Named Entity Recognition API used by TEI Publisher

TEI Publisher Named Entity Recognition API This repository contains the API used by TEI Publisher's web-annotation editor to detect entities in the in

e-editiones.org 14 Nov 15, 2022
Yet Another Sequence Encoder - Encode sequences to vector of vector in python !

Yase Yet Another Sequence Encoder - encode sequences to vector of vectors in python ! Why Yase ? Yase enable you to encode any sequence which can be r

Pierre PACI 12 Aug 19, 2021
Host your own GPT-3 Discord bot

GPT3 Discord Bot Host your own GPT-3 Discord bot i'd host and make the bot invitable myself, however GPT3 terms of service prohibit public use of GPT3

[something hillarious here] 8 Jan 07, 2023
Translation for Trilium Notes. Trilium Notes 中文版.

Trilium Translation 中文说明 This repo provides a translation for the awesome Trilium Notes. Currently, I have translated Trilium Notes into Chinese. Test

743 Jan 08, 2023
Multi Task Vision and Language

12-in-1: Multi-Task Vision and Language Representation Learning Please cite the following if you use this code. Code and pre-trained models for 12-in-

Meta Research 711 Jan 08, 2023
A flask application to predict the speech emotion of any .wav file.

This is a speech emotion recognition app. It will allow you to train a modular MLP model with the RAVDESS dataset, and then use that model with a flask application to predict the speech emotion of an

Aryan Vijaywargia 2 Dec 15, 2021
Create a machine learning model which will predict if the mortgage will be approved or not based on 5 variables

Mortgage-Application-Analysis Create a machine learning model which will predict if the mortgage will be approved or not based on 5 variables: age, in

1 Jan 29, 2022
基于Transformer的单模型、多尺度的VAE模型

UniVAE 基于Transformer的单模型、多尺度的VAE模型 介绍 https://kexue.fm/archives/8475 依赖 需要大于0.10.6版本的bert4keras(当前还没有推到pypi上,可以直接从GitHub上clone最新版)。 引用 @misc{univae,

苏剑林(Jianlin Su) 49 Aug 24, 2022
Automatic privilege escalation for misconfigured capabilities, sudo and suid binaries

GTFONow Automatic privilege escalation for misconfigured capabilities, sudo and suid binaries. Features Automatically escalate privileges using miscon

101 Jan 03, 2023
Source code for CsiNet and CRNet using Fully Connected Layer-Shared feedback architecture.

FCS-applications Source code for CsiNet and CRNet using the Fully Connected Layer-Shared feedback architecture. Introduction This repository contains

Boyuan Zhang 4 Oct 07, 2022
The source code of HeCo

HeCo This repo is for source code of KDD 2021 paper "Self-supervised Heterogeneous Graph Neural Network with Co-contrastive Learning". Paper Link: htt

Nian Liu 106 Dec 27, 2022
Use fastai-v2 with HuggingFace's pretrained transformers

FastHugs Use fastai v2 with HuggingFace's pretrained transformers, see the notebooks below depending on your task: Text classification: fasthugs_seq_c

Morgan McGuire 111 Nov 16, 2022