Wav2vec2 Xlsr 300m Finnish Lm

🚀 Wav2Vec2 XLS-R for Finnish ASR

This acoustic model is a fine - tuned version of facebook/wav2vec2-xls-r-300m for Finnish Automatic Speech Recognition (ASR). It has been fine - tuned using 275.6 hours of Finnish transcribed speech data. Wav2Vec2 XLS - R was introduced in this paper and first released at this page.

This repository also includes a Finnish KenLM language model used in the decoding phase with the acoustic model.

Note: this model is identical to the Finnish - NLP/wav2vec2-xlsr-300m-finnish-lm model. It has simply been copied/moved to the Finnish - NLP Hugging Face organization.

✨ Features

• Multilingual Pretraining: Based on Facebook AI's large - scale multilingual pretrained model Wav2Vec2 XLS - R, pretrained on 436k hours of unlabeled speech across 128 languages.
• Finnish ASR Fine - tuning: Fine - tuned specifically for Finnish ASR using 275.6 hours of Finnish transcribed speech data.
• Included Language Model: Comes with a Finnish KenLM language model for the decoding phase.

📦 Installation

No installation steps are provided in the original README, so this section is skipped.

💻 Usage Examples

Check the run - finnish - asr - models.ipynb notebook in this repository for a detailed example on how to use this model.

📚 Documentation

Model description

Wav2Vec2 XLS - R is Facebook AI's large - scale multilingual pretrained model for speech. It is pretrained on 436k hours of unlabeled speech, including VoxPopuli, MLS, CommonVoice, BABEL, and VoxLingua107. It uses the wav2vec 2.0 objective, in 128 languages.

You can read more about the pretrained model from [this blog](https://ai.facebook.com/blog/xls - r - self - supervised - speech - processing - for - 128 - languages) and this paper.

This model is a fine - tuned version of the pretrained model (300 million parameter variant) for Finnish ASR.

Intended uses & limitations

How to use

You can use this model for Finnish ASR (speech - to - text) task. Refer to the run - finnish - asr - models.ipynb notebook for detailed usage.

Limitations and bias

• Audio Length: This model was fine - tuned with audio samples of maximum length 20 seconds, so it likely works best for short audios of similar length. However, you can try it with longer audios too. If you encounter out - of - memory errors with very long audio files, you can use the audio chunking method introduced in [this blog post](https://huggingface.co/blog/asr - chunking).
• Data Domain: A vast majority of the data used for fine - tuning was from the Finnish Parliament dataset. So, this model may not generalize well to very different domains like common daily spoken Finnish with dialects. Also, the datasets' audios tend to be adult - male dominated, so the model may not work as well for children and women's speeches.
• Language Model: The Finnish KenLM language model used in the decoding phase was trained with text data from audio transcriptions and a subset of Finnish Wikipedia. Thus, it may not generalize to very different language, such as spoken daily language with dialects. It may be beneficial to train your own KenLM language model for your domain language and use it in the decoding.

Training data

This model was fine - tuned with 275.6 hours of Finnish transcribed speech data from the following datasets:

Datasets were filtered to include audio samples with a maximum length of 20 seconds.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e - 04
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: 8 - bit Adam with betas=(0.9, 0.999) and epsilon = 1e - 08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

The pretrained facebook/wav2vec2 - xls - r - 300m model was initialized with the following hyperparameters:

• attention_dropout: 0.094
• hidden_dropout: 0.047
• feat_proj_dropout: 0.04
• mask_time_prob: 0.082
• layerdrop: 0.041
• activation_dropout: 0.055
• ctc_loss_reduction: "mean"

Training results

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0

Evaluation results

Evaluation was done with the [Common Voice 7.0 Finnish test split](https://huggingface.co/datasets/mozilla - foundation/common_voice_7_0).

To evaluate this model, run the eval.py script in this repository:

python3 eval.py --model_id aapot/wav2vec2 - xlsr - 300m - finnish - lm --dataset mozilla - foundation/common_voice_7_0 --config fi --split test

This model (the third row of the table) achieves the following WER (Word Error Rate) and CER (Character Error Rate) results compared to other models:

🔧 Technical Details

• Model Architecture: Based on the Wav2Vec2 XLS - R architecture, with specific hyperparameters for the pretrained and fine - tuned models.
• Training Process: Fine - tuned on a Tesla V100 GPU using Hugging Face's provided training script, with custom data loading for the specific datasets.
• Language Model Training: The Finnish KenLM language model was trained with text data from audio transcriptions and a subset of Finnish Wikipedia.

📄 License

The model is licensed under the Apache - 2.0 license.

Team Members

• Aapo Tanskanen, Hugging Face profile, LinkedIn profile
• Rasmus Toivanen, Hugging Face profile, LinkedIn profile

Feel free to contact us for more details 🤗