Sámi Wav2vec2-Large
The large model pre-trained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz.
Note: This model does not have a tokenizer as it was pre-trained on audio alone. In order to use this model
speech recognition, a tokenizer should be created and the model should be fine-tuned on labeled text data. Check out
this blog for a more in-depth explanation of how to fine-tune the model.
Model description
The Sámi Wav2Vec2 Large has the same architecture and uses the same training objective as the English and multilingual one described in
Paper. It is pre-trained on 22.4k hours of unlabeled Sámi speech from
KAVI radio and television archive materials.
You can read more about the pre-trained model from
this paper.
Intended uses & limitations
You can use this model for Sámi ASR (speech-to-text) and SER (Spoken Emotion Recognition) tasks.
How to use
See
this notebook for more information on how to fine-tune the model.
Limitations and bias
This model was pre-trained with audio samples whose maximum length was 30 seconds so this model most likely works the best for quite short audios of similar length. However, you can try this model with a lot longer audios too and see how it works. If you encounter out-of-memory errors with very long audio files you can use the audio chunking method introduced in
this blog post.
The data used for pre-training was from the
KAVI archives so this model might have biases towards the voices of radio hosts. The pre-training data was filtered via neural
VAD, but some non-speech events like music might be still present in the training data, which might cause issues when fine-tuned on clear (no background noise) speech.
Training data
This model was pre-trained with 22.4k hours of Sámi speech data from the following sources:
Datasets were filtered to include a maximum length of 30 seconds long audio samples.
Training procedure
Training was done on 256 AMD MI250x GPU modules (512 GPUs from the software perspective), using
LUMI.
The training script was provided by Fairseq and it is available
here.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 4e-04
- max_update: 166667
- seed: 1
- optimizer: 8-bit Adam with betas=(0.9,0.98) and epsilon=1e-06
- lr_scheduler_type: linear
- lr_scheduler_warmup_updates: 5000
- fp16: true
- max_sample_size: 960000
- min_sample_size: 32000
- normalize: true
- max_tokens: 1800000
- distributed_world_size: 512
The pre-trained model was initialized with the following hyperparameters:
- quantize_targets: true
- latent_temp: [2.0, 0.5, 0.999995]
- extractor_mode: layer_norm
- layer_norm_first: true
- dropout_input: 0.0
- dropout_features: 0.0
- feature_grad_mult: 1.0
- encoder_embed_dim: 1024
- encoder_layers: 24
- encoder_ffn_embed_dim: 4096
- encoder_attention_heads: 16
- dropout: 0.0
- attention_dropout: 0.0
- activation_dropout: 0.0
- encoder_layerdrop: 0.0
Training results
| Training Loss | Epoch | Validation Loss |
|---|
| 4.909 | 1 | 4.707 |
| 2.452 | 25 | 2.356 |
| 2.245 | 50 | 2.11 |
| 2.094 | 75 | 1.985 |
| 2.000 | 100 | 1.916 |
Framework versions
- Pytorch 2.3.0+rocm6.0
- Fairseq 0.12.2
Citation
If you use our models or scripts, please cite our article as:
1@inproceedings{getman-etal-2025-towards,
2 title = "Towards large-scale speech foundation models for a low-resource minority language",
3 author = "Getman, Yaroslav and
4 Gr{\'o}sz, Tam{\'a}s and
5 Hiovain-Asikainen, Katri and
6 Lehtonen, Tommi and
7 Kurimo, Mikko",
8 editor = "Johansson, Richard and
9 Stymne, Sara",
10 booktitle = "Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies (NoDaLiDa/Baltic-HLT 2025)",
11 month = mar,
12 year = "2025",
13 address = "Tallinn, Estonia",
14 publisher = "University of Tartu Library",
15 url = "https://aclanthology.org/2025.nodalida-1.19/",
16 pages = "192--200",
17 ISBN = "978-9908-53-109-0"
18}
Team Members
Feel free to contact us for more details 🤗
