README.md

metadata

license: cc-by-nc-4.0
language:
  - ab
  - af
  - am
  - ar
  - as
  - az
  - ba
  - be
  - bn
  - bo
  - bs
  - br
  - bg
  - ca
  - cs
  - cv
  - cy
  - da
  - de
  - dv
  - el
  - en
  - eo
  - et
  - eu
  - ee
  - fo
  - fa
  - tl
  - fi
  - fr
  - fy
  - ga
  - gl
  - gv
  - gn
  - gu
  - ht
  - ha
  - he
  - hi
  - hr
  - hu
  - hy
  - ig
  - ia
  - id
  - is
  - it
  - jv
  - ja
  - kn
  - ka
  - kk
  - km
  - rw
  - ky
  - ku
  - ko
  - lo
  - la
  - lv
  - ln
  - lt
  - lb
  - lg
  - ml
  - mr
  - mk
  - mg
  - mt
  - mn
  - mi
  - ms
  - my
  - ne
  - nl
  - nn
  - 'no'
  - oc
  - or
  - pa
  - pl
  - pt
  - ps
  - ro
  - ru
  - sa
  - si
  - sl
  - sk
  - sn
  - sd
  - so
  - st
  - es
  - sq
  - sc
  - sr
  - su
  - sw
  - sv
  - ta
  - tt
  - te
  - tg
  - th
  - tn
  - tk
  - tr
  - tw
  - ug
  - uk
  - ur
  - uz
  - vi
  - xh
  - yi
  - yo
  - zh

Table of Contents:

1. Summary
2. Training Data and Code
3. ML-SUPERB Scores
4. Languages and Datasets
5. Citing and Funding Information

mHuBERT-147 models

mHuBERT-147 are compact and competitive multilingual HuBERT models trained on 90K hours of open-license data in 147 languages. Different from traditional HuBERTs, mHuBERT-147 models are trained using faiss IVF discrete speech units. Training employs a two-level language, data source up-sampling during training. See more information in our paper.

This repository contains:

• Fairseq checkpoint (original);
• HuggingFace checkpoint;
• Faiss index for continuous pre-training (OPQ16_64,IVF1000_HNSW32,PQ16x4fsr).

Model details: 3rd iteration, base architecture, 147 languages.

Training

Manifest list: https://huggingface.co/utter-project/mHuBERT-147-base-3rd-iter/tree/main/manifest

Please note that since training, there were CommonVoice removal requests. This means that some of the listed files are no longer available.

Fairseq fork: https://github.com/utter-project/fairseq

Scripts for pre-processing/faiss clustering: https://github.com/utter-project/mHuBERT-147-scripts

ML-SUPERB Scores

Languages and Datasets

For ASR/ST/TTS datasets, only train set is used.

• Aishell and AISHELL-3
• BibleTTS
• ClovaCall
• CommonVoice v11
• Google TTS data: Javanese, Khmer, Nepali, Sundanese, South African Languages, Bengali Languages
• IISc-MILE: Tamil, Kannada
• Japanese Versatile Speech
• Kokoro
• Kosp2e
• Media Speech: Turkish Only
• Multilingual LibriSpeech
• Samrómur
• THCHS-30 and THUYG-20
• VoxLingua107
• VoxPopuli

Languages present not indexed by Huggingface: Asturian (ast), Basaa (bas), Cebuano (ceb), Central Kurdish/Sorani (ckb), Hakha Chin (cnh), Hawaiian (haw), Upper Sorbian (hsb) Kabyle (kab), Moksha (mdf), Meadow Mari (mhr), Hill Mari (mrj), Erzya (myv), Taiwanese Hokkien (nan-tw), Sursilvan (rm-sursilv), Vallader (rm-vallader), Sakha (sah), Santali (sat), Scots (sco), Saraiki (skr), Tigre (tig), Tok Pisin (tpi), Akwapen Twi (tw-akuapem), Asante Twi (tw-asante), Votic (vot), Waray (war), Cantonese (yue).

Citing and Funding Information

@inproceedings{boito2024mhubert,
author={Marcely Zanon Boito, Vivek Iyer, Nikolaos Lagos, Laurent Besacier, Ioan Calapodescu},
title={{mHuBERT-147: A Compact Multilingual HuBERT Model}},
year=2024,
booktitle={Interspeech 2024},
}

This is an output of the European Project UTTER (Unified Transcription and Translation for Extended Reality) funded by European Union’s Horizon Europe Research and Innovation programme under grant agreement number 101070631.

For more information please visit https://he-utter.eu/