yamine15 / asr2k

asr2k

ASR2K

This repository will contain our code of our publication at interspeech 2022

Li, Xinjian, et al. "ASR2K: Speech Recognition for Around 2000 Languages without Audio" Interspeech 2022. 2022

We plan to release ASR models for 2k languages (currently 1909 languages). The architecture is as follows:

It has a modular architecture instead of modern end-to-end models

As most of the languages do not have any training set (and even test set), it is impossible to create any end-to-end models directly. Instead of building the end-to-end models, We are interested in decomposing the entire model into a sequence of linguistic components. it has a few advantages in this scenario:

• some of the components here are relatively less dependent on training resources (e.g: texts are easier to obtain than audios)
• some of them are already well-defined by linguists (e.g: phonetics, phonology), those domain knowledge can be incorporated into the model through some Bayesian frameworks.
• language-wise interpolation is easier sometimes (e.g: English and German shares many common phones)

Model

Our model mainly consists of three components: acoustic model, pronunciation model and language model. The acoustic model and pronunciation model have been released separately on Github:

• acoustic model: allosaurus
• pronunciation model: transphone

The language model is based on icefall recipes and crubadan database. We are currently working on releasing the language model and organize all codes into a simple codebase in this repository