Efficient Speech Translation with Dynamic Latent Perceivers

This work was accepted at ICASSP 2023 and is availble here.

Transformers have been the dominant architecture for Speech Translation in recent years, achieving significant improvements in translation quality. Since speech signals are longer than their textual counterparts, and due to the quadratic complexity of the Transformer, a down-sampling step is essential for its adoption in Speech Translation. Instead, in this research, we propose to ease the complexity by using a Perceiver encoder to map the speech inputs to a fixed-length latent representation. Furthermore, we introduce a novel way of training Perceivers, with Dynamic Latent Access (DLA), unlocking larger latent spaces without any additional computational overhead. Speech-to-Text Perceivers with DLA can match the performance of Transformer baselines across three language pairs in MuST-C. Finally, a DLA-trained model is easily adaptable to DLA at inference, and can be flexibly deployed with various computational budgets, without significant drops in translation quality.

S2T-Perceiver	Dynamic Latent Access

Setup

# path to clone this repo
export PERCEIVER_ROOT=...
# path to must-c data
export MUSTC_ROOT=...
# path to save experiment outputs
export OUTPUT_ROOT=...

Clone this repository to $PERCEIVER_ROOT:

git clone https://github.com/mt-upc/s2t_perceiver.git ${PERCEIVER_ROOT} && \
cd ${PERCEIVER_ROOT} && \
git submodule init && \
git submodule update

Create a conda environment using the environment.yml file and activate it:

conda env create -f ${PERCEIVER_ROOT}/environment.yml && \
conda activate s2t_perceiver && \
pip install --editable ${PERCEIVER_ROOT}/fairseq/

To prepare the MuST-C data follow the instructions here. We used en-de (v2.0) and en-es, en-ru (v1.0).

Train an S2T-Perceiver with Dynamic Latent Access (DLA-train)

To train the model first do the ASR pre-training step and then start the ST training with the pre-trained encoder.

• To train without DLA-train, set $k_train = $n.
• Example is for English-to-German (en-de).
• The suggested values for base_update_freq and batch_size are for an NVIDIA GeForce RTX 2080 Ti , adjust them accordingly for other devices.
• For training with multiple devices make sure that base_update_freq is divisible by n_gpus.

# total number of latents
n=...
# number of latents for DLA-train
k_train=...
# language pair, for example: "en-de"
lang_pair=...

# ASR pre-training
bash ${PERCEIVER_ROOT}/scripts/train_perceiver_asr.sh $n $k_train

# ST training
bash ${PERCEIVER_ROOT}/scripts/train_perceiver_st.sh $lang_pair $n $k_train

Evaluate an S2T-Perceiver with Dynamic Latent Access (DLA-inf)

To evaluate without DLA-inf, set $k_inf = $n.

# path to the trained model ($path_to_exp/ckpts)
path_to_exp=...
# language pair, for example: "en-de"
lang_pair=...
# number of latents for DLA-inf
k_inf=...

bash ${PERCEIVER_ROOT}/scripts/eval_perceiver_st.sh $path_to_exp $lang_pair $k_inf

Citation

@INPROCEEDINGS{10095276,
  author={Tsiamas, Ioannis and Gállego, Gerard I. and Fonollosa, José A. R. and Costa-jussà, Marta R.},
  booktitle={ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={{Efficient Speech Translation with Dynamic Latent Perceivers}}, 
  year={2023},
  volume={},
  number={},
  pages={1-5},
  doi={10.1109/ICASSP49357.2023.10095276}}