Røst-315m

This is a Danish state-of-the-art speech recognition model, trained by the Alexandra Institute.

Quick Start

Start by installing the required libraries:

$ pip install transformers kenlm pyctcdecode

Next you can use the model using the transformers Python package as follows:

>>> from transformers import pipeline
>>> audio = get_audio()  # 16kHz raw audio array
>>> transcriber = pipeline(model="alexandrainst/roest-315m")
>>> transcriber(audio)
{'text': 'your transcription'}

Evaluation Results

We have evaluated both our and existing models on the CoRal test set as well as the Danish Common Voice 17 test set. To ensure as robust an evaluation as possible, we have bootstrapped the results 1000 times and report here the mean scores along with a 95% confidence interval (lower is better; best scores in bold, second-best in italics):

Model	Number of parameters	CoRal CER	CoRal WER	Danish Common Voice 17 CER	Danish Common Voice 17 WER
Røst-315m (this model)	315M	6.9% ± 0.2%	14.9% ± 0.4%	5.1% ± 0.6%	13.2% ± 0.8%
chcaa/xls-r-300m-danish-nst-cv9	315M	14.4% ± 0.3%	36.5% ± 0.6%	4.1% ± 0.5%	12.0% ± 0.8%
mhenrichsen/hviske	1540M	15.8% ± 0.7%	36.5% ± 1.0%	5.3% ± 0.4%	14.5% ± 0.8%
openai/whisper-large-v3	1540M	16.5% ± 1.3%	36.8% ± 1.9%	7.6% ± 0.6%	18.3% ± 1.1%
openai/whisper-large-v2	1540M	19.7% ± 1.8%	42.2% ± 2.6%	10.6% ± 1.6%	23.3% ± 2.0%
openai/whisper-large	1540M	19.5% ± 1.3%	42.4% ± 1.7%	12.8% ± 0.8%	28.3% ± 1.3%
openai/whisper-medium	764M	21.5% ± 1.7%	47.4% ± 2.6%	13.3% ± 0.8%	30.0% ± 1.3%
openai/whisper-small	242M	26.1% ± 1.2%	57.9% ± 1.5%	22.9% ± 4.3%	49.3% ± 6.3%
openai/whisper-base	73M	50.8% ± 3.6%	100.1% ± 5.6%	43.1% ± 5.0%	85.1% ± 7.9%
openai/whisper-tiny	38M	63.7% ± 3.9%	120.3% ± 5.7%	58.5% ± 5.8%	106.4% ± 8.7%

Detailed Evaluation Across Demographics on the CoRal Test Set

Dialect	CER	WER
Københavnsk	2.8%	6.6%
Sjællandsk	3.9%	8.6%
Fynsk	7.3%	15.7%
Sønderjysk	13.3%	27.1%
Vestjysk	11.1%	24.9%
Østjysk	3.2%	7.6%
Nordjysk	2.7%	5.7%
Sydømål	6.0%	12.3%
Bornholmsk	9.1%	19.9%
Non-native	7.3%	16.7%

Gender	CER	WER
Female	8.0%	17.1%
Male	5.8%	12.8%

Age group	CER	WER
0-25	5.5%	12.1%
25-50	5.9%	13.3%
50+	8.2%	17.5%

Training Data

This model is the result of four different stages of training:

"Pretraining" on 436,000 hours of unlabelled multilingual publicly available data, 13,628 hours of which is Danish. Pretraining here means that the model learnt to "fill in" gaps of raw audio - no transcriptions were used (or available) during this process. The pretraining data is distributed as follows:
- 372,000 hours from VoxPopuli, being speeches from the European Parliament in 23 European languages. This includes 13,600 hours of Danish speech.
- 51,000 hours from Multilingual LibriSpeech, being audiobooks in 8 European languages. This does not include any Danish speech.
- 7,000 hours from Common Voice 6, being read-aloud speech in 60 diverse languages. This does not include any Danish speech.
- 6,600 hours from VoxLingua107, being audio from YouTube videos in 107 languages. This includes 28 hours of Danish speech.
- 1,000 hours from BABEL, being conversational telephone speech in 17 African and Asian languages. This does not include any Danish speech.
Continued pretraining on 141,000 hours of Danish radio (more specifically, DR P1 and Radio24Syv from 2005 to 2021).
"Finetuning" on 373 hours of labelled Danish publicly available data. "Finetuning" indicates that this stage of training was supervised, i.e. the model was trained on both audio and transcriptions to perform the speech-to-text task (also known as automatic speech recognition). The finetuning data is as follows:
- The read-aloud training split of the CoRal dataset (revision fb20199b3966d3373e0d3a5ded2c5920c70de99c), consisting of 361 hours of Danish read-aloud speech, diverse across dialects, accents, ages and genders.
- The Danish training split of the Common Voice 17 dataset, consisting of 12 hours of Danish read-aloud speech.
An n-gram language model has been trained separately, and is used to guide the transcription generation of the finetuned speech recognition model. This n-gram language model has been trained on all of the Danish Wikipedia (approximately 287,000 articles).

The first step was trained by Babu et al. (2021), second step by Hansen (2022) and the third and fourth step by Nielsen et al. (2024).

The final product is then the combination of the finetuned model along with the n-gram model, and this is what is used when you use the model as mentioned in the Quick Start section above.

Intended use cases

This model is intended to be used for Danish automatic speech recognition.

Note that Biometric Identification is not allowed using the CoRal dataset and/or derived models. For more information, see addition 4 in our license.

Why the name Røst?

Røst is both the Danish word for the human voice as well as being the name of one of the cold-water coral reefs in Scandinavia.

License

The dataset is licensed under a custom license, adapted from OpenRAIL-M, which allows commercial use with a few restrictions (speech synthesis and biometric identification). See license.

Creators and Funders

The CoRal project is funded by the Danish Innovation Fund and consists of the following partners:

Citation

We will submit a research paper soon, but until then, if you use this model in your research or development, please cite it as follows:

@dataset{coral2024,
  author    = {Dan Saattrup Nielsen, Sif Bernstorff Lehmann, Simon Leminen Madsen, Anders Jess Pedersen, Anna Katrine van Zee, Anders Søgaard and Torben Blach},
  title     = {CoRal: A Diverse Danish ASR Dataset Covering Dialects, Accents, Genders, and Age Groups},
  year      = {2024},
  url       = {https://hf.co/datasets/alexandrainst/coral},
}

alexandrainst
/

roest-315m