Finnish Wav2vec2-XLarge

The x-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 more in-detail explanation of how to fine-tune the model.

Note: Fine-tuned version is available at GetmanY1/wav2vec2-xlarge-fi-150k-finetuned

Model description

The Finnish Wav2Vec2 X-Large has the same architecture and uses the same training objective as the multilingual one described in paper. It is pre-trained on 158k hours of unlabeled Finnish speech, including KAVI radio and television archive materials, Lahjoita puhetta (Donate Speech), Finnish Parliament, Finnish VoxPopuli.

You can read more about the pre-trained model from this paper. The training scripts are available on GitHub.

Intended uses & limitations

You can use this model for Finnish 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 60 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.

A vast majority of the data used for pre-training was from the KAVI archives so this model might have biases towards the voices of tv and radio hosts, as well as to colloquial Finnish. The pre-training data was filtered via neural VAD, but some non-speech events like music might be still present in the traning data, which might cause issues when fine-tuned on clear (no background noise) speech.

Training data

This model was pre-trained with 158k hours of Finnish speech data from the following sources:

Datasets were filtered to include a maximum length of 60 seconds long audio samples.

Training procedure

Training was done on 256 AMD MI250x GPU modules (512 GPUs from the software perspective), using LUMI, during the Second Finnish LUMI Extreme Scale.

Training script was provided by Fairseq and it is available here.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-04
• max_update: 390625
• seed: 1
• optimizer: 8-bit Adam with betas=(0.9,0.98) and epsilon=1e-06
• lr_scheduler_type: linear
• lr_scheduler_warmup_updates: 12500
• fp16: true
• max_sample_size: 960000
• min_sample_size: 32000
• normalize: true
• max_tokens: 960000
• 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: 0.1
• dropout_features: 0.1
• feature_grad_mult: 1.0
• encoder_embed_dim: 1280
• encoder_layers: 48
• encoder_ffn_embed_dim: 5120
• encoder_attention_heads: 16
• dropout: 0.0
• attention_dropout: 0.0
• encoder_layerdrop: 0.0

Training results

Framework versions

• Pytorch 1.13.1+rocm5.2
• Fairseq 0.12.2

Citation

If you use our models or scripts, please cite our article as:

@inproceedings{getman25_interspeech,
  title     = {{Is your model big enough? Training and interpreting large-scale monolingual speech foundation models}},
  author    = {{Yaroslav Getman and Tamás Grósz and Tommi Lehtonen and Mikko Kurimo}},
  year      = {{2025}},
  booktitle = {{Interspeech 2025}},
  pages     = {{231--235}},
  doi       = {{10.21437/Interspeech.2025-46}},
  issn      = {{2958-1796}},
}

Team Members

• Yaroslav Getman, Hugging Face profile, LinkedIn profile
• Tamas Grosz, Hugging Face profile, LinkedIn profile

Feel free to contact us for more details 🤗