jamfly / allost

Official Implementation of AlloST: Low-resource Speech Translation without Source Transcription

This repository includes the implementation of the INTERSPEECH 2021 paper:

AlloST: Low-resource Speech Translation without Source Transcription

Dependency:

1. ESPnet: please follow the instruction to install ESPnet
2. Allosaurus: pip install allosaurus

How to run:

1. Install dependencies
2. cd to egs/fisher_callhome/st1
3. bash the_experiment_you_want_to_run

Results

Method	fisher-dev	fisher-dev2	fisher-test
Conformer	20.84	21.49	21.06
Conformer w/ MTL	31.21	32.59	31.47
Phoneme Segment	21.00	N/A	19.8
Phoneme Concat.	34.50	N/A	33.00
Word Embedding	23.45	N/A	21.72
Encoder Fusion	29.52	30.01	29.01
Decoder Fusion	26.32	27.14	26.1
Encoder Fusion + Decoder Fusion w/o BPE	26.52	26.92	26.12
Encoder Fusion + Decoder Fusion w/ BPE 1k	28.95	30.18	29.49
Encoder Fusion + Decoder Fusion w/ BPE 10k	29.26	29.90	29.37
Encoder Fusion + Decoder Fusion w/ BPE 16k	28.89	30.18	29.73
Encoder Fusion + Decoder Fusion w/ BPE 32k	29.19	31.05	30.30
Encoder Fusion + Decoder Fusion w/ BPE 48k	29.37	30.26	29.70
Encoder Fusion + Decoder Fusion w/ BPE 64k	29.44	30.73	29.12

Citation:

If you want to publish experimental results with our idea, please cite the following articles:

@misc{2105.00171,
Author = {Yao-Fei Cheng and Hung-Shin Lee and Hsin-Min Wang},
Title = {AlloST: Low-resource Speech Translation without Source Transcription},
Year = {2021},
Eprint = {arXiv:2105.00171},
}