--- license: apache-2.0 language: fi tags: - automatic-speech-recognition - fi - finnish model-index: - name: wav2vec2-large-uralic-voxpopuli-v2-1500h results: - task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Lahjoita puhetta (Donate Speech) type: lahjoita-puhetta args: fi metrics: - name: Dev WER type: wer value: 19.14 - name: Dev CER type: cer value: 5.05 - name: Test WER type: wer value: 20.49 - name: Test CER type: cer value: 5.93 --- # Colloquial Finnish Wav2vec2-Large ASR [facebook/wav2vec2-large-uralic-voxpopuli-v2](https://huggingface.co/facebook/wav2vec2-large-uralic-voxpopuli-v2) fine-tuned on 1500 hours of [Lahjoita puhetta (Donate Speech)](https://link.springer.com/article/10.1007/s10579-022-09606-3) on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. ## Model description The Finnish Wav2Vec2 Large has the same architecture and uses the same training objective as the English and multilingual one described in [Paper](https://arxiv.org/abs/2006.11477). You can read more about the pre-trained model from [this paper](https://www.isca-archive.org/interspeech_2024/getman24_interspeech.html). The training scripts are available on [GitHub](https://github.com/aalto-speech/colloquial-Finnish-wav2vec2) ## Intended uses & limitations You can use this model for Finnish ASR (speech-to-text). ### How to use To transcribe audio files the model can be used as a standalone acoustic model as follows: ``` from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC from datasets import load_dataset import torch # load model and processor processor = Wav2Vec2Processor.from_pretrained("GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-1500h") model = Wav2Vec2ForCTC.from_pretrained("GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-1500h") # load dummy dataset and read soundfiles ds = load_dataset("mozilla-foundation/common_voice_16_1", "fi", split='test') # tokenize input_values = processor(ds[0]["audio"]["array"], return_tensors="pt", padding="longest").input_values # Batch size 1 # retrieve logits logits = model(input_values).logits # take argmax and decode predicted_ids = torch.argmax(logits, dim=-1) transcription = processor.batch_decode(predicted_ids) ``` ### Limitations and bias This model was fine-tuned with audio samples whose maximum length was 50 seconds so this model most likely works the best for 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](https://huggingface.co/blog/asr-chunking). The model was fine-tuned on the data from the [Lahjoita puhetta (Donate Speech) corpus](https://link.springer.com/article/10.1007/s10579-022-09606-3) so this model might have biases towards colloquial Finnish. ## Evaluation results Evaluation results in terms of WER (word error rate) and CER (character error rate) on the Lahjoita puhetta dev and test sets: |**System** | **Labeled training data, h** | **dev WER [%]** | **dev CER [%]** | **test WER [%]** | **test CER [%]** | |-----------------------------------------------------------------------------------------------------------------------------|------------------------------|-------------|--------------|---------------|---------------| |**Base Models** |[GetmanY1/wav2vec2-base-fi-voxpopuli-v2-100h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-voxpopuli-v2-100h) |100 |29.35|7.94|31.90|9.05| |[GetmanY1/wav2vec2-base-fi-voxpopuli-v2-1500h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-voxpopuli-v2-100h) |1500|22.18|5.96|24.43|6.97| |[GetmanY1/wav2vec2-base-fi-lp-from-scratch-100h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-lp-from-scratch-100h) |100 |26.40|6.86|28.92|8.09| |[GetmanY1/wav2vec2-base-fi-lp-from-scratch-1500h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-lp-from-scratch-1500h) |1500|21.61|5.59|24.35|6.87| |[GetmanY1/wav2vec2-base-fi-lp-cont-pt-100h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-lp-cont-pt-100h) |100 |22.49|5.84|24.95|7.09| |[GetmanY1/wav2vec2-base-fi-lp-cont-pt-1500h](https://huggingface.co/GetmanY1/wav2vec2-base-fi-lp-cont-pt-1500h) |1500|**17.38**|**4.61**|**19.65**|**5.69**| |**Large Models** |[GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-100h](https://huggingface.co/GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-100h) |100 |21.02|5.70|22.98|6.90| |[GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-1500h](https://huggingface.co/GetmanY1/wav2vec2-large-uralic-voxpopuli-v2-1500h)|1500|19.14|5.05|20.49|5.93| |[GetmanY1/wav2vec2-large-fi-lp-from-scratch-100h](https://huggingface.co/GetmanY1/wav2vec2-large-fi-lp-from-scratch-100h) |100 |21.66|5.61|23.85|6.76| |[GetmanY1/wav2vec2-large-fi-lp-from-scratch-1500h](https://huggingface.co/GetmanY1/wav2vec2-large-fi-lp-from-scratch-1500h) |1500|17.54|4.59|19.26|5.58| |[GetmanY1/wav2vec2-large-fi-lp-cont-pt-100h](https://huggingface.co/GetmanY1/wav2vec2-large-fi-lp-cont-pt-100h) |100 |20.20|5.40|22.81|6.64| |[GetmanY1/wav2vec2-large-fi-lp-cont-pt-1500h](https://huggingface.co/GetmanY1/wav2vec2-large-fi-lp-cont-pt-1500h) |1500|**16.24**|**4.34**|**18.04**|**5.29**| ## Citation If you use our models or scripts, please cite our article as: ```bibtex @inproceedings{getman24_interspeech, title = {What happens in continued pre-training? Analysis of self-supervised speech models with continued pre-training for colloquial Finnish ASR}, author = {Yaroslav Getman and Tamas Grosz and Mikko Kurimo}, year = {2024}, booktitle = {Interspeech 2024}, pages = {5043--5047}, doi = {10.21437/Interspeech.2024-476}, } ``` ## Team Members - Yaroslav Getman, [Hugging Face profile](https://huggingface.co/GetmanY1), [LinkedIn profile](https://www.linkedin.com/in/yaroslav-getman/) - Tamas Grosz, [Hugging Face profile](https://huggingface.co/Grosy), [LinkedIn profile](https://www.linkedin.com/in/tam%C3%A1s-gr%C3%B3sz-950a049a/) Feel free to contact us for more details 🤗