Orange/SSA-HuBERT-base-60k-V2

Model description

This self-supervised speech model (a.k.a. SSA-HuBERT-base-60k-V2) 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.

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			WER
	base-V2	large	XL	base-V2	large	XL
Afrikaans	19.8	13.0	12.4	59.1	42.3	39.8
Amharic	13.3	9.9	10.3	44.3	32.9	34.3
Fula	16.8	15.4	16.4	54.2	50.9	52.7
Ganda	10.3	9.4	9.0	49.4	46.9	45.6
Hausa	8.5	6.6	5.5	28.1	21.6	19.6
Igbo	15.8	13.2	12.8	49.7	44.2	43.3
Kamba	14.5	11.4	10.7	50.2	41.8	39.7
Lingala	6.9	4.9	4.3	20.4	14.9	13.6
Luo	7.6	6.1	5.8	33.6	28.0	27.0
Northen-Sotho	10.7	8.4	8.0	35.9	28.8	33.7
Nyanja	10.6	8.0	7.0	44.5	35.3	32.7
Oromo	19.4	18.2	18.3	73.1	66.9	67.7
Shona	7.3	5.1	4.7	34.6	24.6	23.2
Somali	19.1	15.5	15.3	58.6	49.8	49.2
Swahili	4.8	3.3	2.7	17.6	12.0	10.1
Umbundu	18.3	15.1	14.6	53.7	47.7	50.6
Wolof	16.3	13.7	12.4	48.7	42.2	40.0
Xhosa	8.9	6.7	6.3	42.2	34.6	33.5
Yoruba	21.6	19.9	19.0	62.2	57.9	55.9
Zulu	9.1	6.7	6.2	42.1	33.3	31.0
Overall average	13.0	10.5	10.1	45.1	37.8	37.2

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.