Orange/SSA-HuBERT-base-60k

Model description

This self-supervised speech model (a.k.a. SSA-HuBERT-base-60k) is based on a HuBERT Base architecture (~95M params) [1].
It was trained on nearly 60 000 hours of speech segments and covers 21 languages and variants spoken in Sub-Saharan Africa.

Pretraining data

Dataset: The training dataset was composed of both studio recordings (controlled environment, prepared talks) and street interviews (noisy environment, spontaneous speech).
Languages: Bambara (bam), Dyula (dyu), French (fra), Fula (ful), Fulfulde (ffm), Fulfulde (fuh), Gulmancema (gux), Hausa (hau), Kinyarwanda (kin), Kituba (ktu), Lingala (lin), Luba-Lulua (lua), Mossi (mos), Maninkakan (mwk), Sango (sag), Songhai (son), Swahili (swc), Swahili (swh), Tamasheq (taq), Wolof (wol), Zarma (dje).

ASR fine-tuning

The SpeechBrain toolkit (Ravanelli et al., 2021) is used to fine-tune the model.
Fine-tuning is done for each language using the FLEURS dataset [2].
The pretrained model (SSA-HuBERT-base-60k) is considered as a speech encoder and is fully fine-tuned with two 1024 linear layers and a softmax output at the top.

License

This model is released under the CC-by-NC 4.0 conditions.

Publication

This model were presented at AfricaNLP 2024. The associated paper is available here: Africa-Centric Self-Supervised Pre-Training for Multilingual Speech Representation in a Sub-Saharan Context

Citation

Please cite our paper when using SSA-HuBERT-base-60k model:

Caubrière, A., & Gauthier, E. (2024). Africa-Centric Self-Supervised Pre-Training for Multilingual Speech Representation in a Sub-Saharan Context. In 5th Workshop on African Natural Language Processing (AfricaNLP 2024).

Bibtex citation:

@inproceedings{caubri{\`e}re2024ssaspeechssl,    
    title={Africa-Centric Self-Supervised Pretraining for Multilingual Speech Representation in a Sub-Saharan Context},    
    author={Antoine Caubri{\`e}re and Elodie Gauthier},    
    booktitle={5th Workshop on African Natural Language Processing},    
    year={2024},    
    url={https://openreview.net/forum?id=zLOhcft2E7}}

Results

The following results are obtained in a greedy mode (no language model rescoring).
Character error rates (CERs) and Word error rates (WERs) are given in the table below, on the 20 languages of the SSA subpart of the FLEURS dataset.

Language	CER	CER (joint finetuning)	WER	WER (joint finetuning)
Afrikaans	23.3	20.3	68.4	62.6
Amharic	15.9	14.9	52.7	49.0
Fula	21.2	17.8	61.9	56.4
Ganda	11.5	10.7	52.8	50.3
Hausa	10.5	9.0	32.5	29.4
Igbo	19.7	17.2	57.5	52.9
Kamba	16.1	15.6	53.9	53.7
Lingala	8.7	6.9	24.7	20.9
Luo	9.9	8.2	38.9	34.9
Northen-Sotho	13.5	11.7	43.2	38.9
Nyanja	13.3	10.9	54.2	48.3
Oromo	22.8	20.1	78.1	74.8
Shona	11.6	8.3	50.2	39.3
Somali	21.6	19.7	64.9	60.3
Swahili	7.1	5.5	23.8	20.3
Umbundu	21.7	18.8	61.7	54.2
Wolof	19.4	17.0	55.0	50.7
Xhosa	11.9	9.9	51.6	45.9
Yoruba	24.3	23.5	67.5	65.7
Zulu	12.2	9.6	53.4	44.9
Overall average	15.8	13.8	52.3	47.7

Reproductibilty

We propose a notebook to reproduce the ASR experiments mentioned in our paper. See SB_ASR_FLEURS_finetuning.ipynb.
By using the ASR_FLEURS-swahili_hf.yaml config file, you will be able to run the recipe on Swahili.

References

[1] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. In 2021 IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp.3451–3460, 2021. doi: 10.1109/TASLP.2021.3122291.
[2] Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. Fleurs: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT), pp. 798–805, 2022. doi: 10.1109/SLT54892.2023.10023141.