Model reference · open weights
mHuBERT-147 is an open-weight embedding model from utter-project, listed in the AxForge catalogue. AxForge can bring it up on EU-owned hardware for you on request — with the licence handled where one is required.
About
This repository contains the best mHuBERT-147 pre-trained model. MODEL DETAILS: 3rd iteration, K=1000, HuBERT base architecture (95M parameters), 147 languages. 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. Table of Contents: 1. Summary 2. Training Data and Code 3. ML-SUPERB Scores 4. Languages and Datasets 5. Intermediate Checkpoints 6. Citing and Funding Information This repository contains: Fairseq checkpoint (original); HuggingFace checkpoint (conversion using transformers library); Faiss index for continuous pre-training (OPQ1664,IVF1000HNSW32,PQ16x4fsr). Related Models: 2nd Iteration mHuBERT-147 1st Iteration mHuBERT-147 CommonVoice Prototype (12 languages) Training Manifest list available here. Please note that since training, there were CommonVoice removal requests. This means that some of the listed files are no longer available. Fairseq fork contains the scripts for training with multilingual batching with two-level up-sampling. Scripts for pre-processing/faiss clustering available here. ML-SUPERB Scores mHubert-147 reaches second and first position in the 10min and 1h leaderboards respectively. We achieve new SOTA scores for three LID tasks. See more information in our paper. Languages and Datasets 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 Hokki
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | utter-project |
|---|---|
| Type | Embedding models |
| Parameters (lead) | 94M |
| Variants | 1 |
| Runs with | transformers |
| Released | 2024-03-14 |
| Popularity | 80k downloads / month |
| Likes | 106 |
| Licence | Commercial licence needed |
How it works
Variants
Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.
| Variant | Params | Precision | VRAM | Fits 16 GB | Weights |
|---|---|---|---|---|---|
| mHuBERT-147 | 94M | BF16 | ~0.2 GB | ✓ | Weights ↗ |
Using it via the API
Once AxForge deploys mhubert-147 for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (mhubert-147 below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/embeddings \
-H "Authorization: Bearer $AXFORGE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"mhubert-147","input":"text to embed"}'
Licence
The weights are open but cc-by-nc-sa-4.0 needs a commercial agreement for business use. AxForge can arrange that licence and host the model for you — you pay AxForge, we settle with the model’s maker. Ask us for a quote. Read the licence ↗
Explore