CoRal

Danish ASR and TTS datasets and models, as part of the CoRal project, funded by the Innovation Fund.

---

Developers:

• Anders Jess Pedersen (anders.j.pedersen@alexandra.dk)
• Dan Saattrup Nielsen (dan.nielsen@alexandra.dk)
• Simon Leminen Madsen (simon.leminen@alexandra.dk)

Installation

1. Run make install, which installs Poetry (if it isn't already installed), sets up a virtual environment and all Python dependencies therein.
2. Run source .venv/bin/activate to activate the virtual environment.
3. Run make to see a list of available commands.

Usage

Finetuning an Acoustic Model for Automatic Speech Recognition (ASR)

You can use the finetune_asr_model script to finetune your own ASR model:

python src/scripts/finetune_asr_model.py [key=value]...

Here are some of the more important available keys:

• model: The base model to finetune. Supports the following values:
  • wav2vec2-small
  • wav2vec2-medium
  • wav2vec2-large
  • whisper-xxsmall
  • whisper-xsmall
  • whisper-small
  • whisper-medium
  • whisper-large
  • whisper-large-turbo
• datasets: The datasets to finetune the models on. Can be a single dataset or an array of datasets (written like [dataset1,dataset2,...]). Supports the following values:
  • coral
  • common_voice_17
  • common_voice_9
  • fleurs
  • ftspeech
  • nota
  • nst
• dataset_probabilities: In case you are finetuning on several datasets, you need to specify the probability of sampling each one. This is an array of probabilities that need to sum to 1. If not set, the datasets are sampled uniformly.
• model_id: The model ID of the finetuned model. Defaults to the model type along with a timestamp.
• push_to_hub, hub_organisation and private: Whether to push the finetuned model to the Hugging Face Hub, and if so, which organisation to push it to. If private is set to True, the model will be private. The default is not to push the model to the Hub.
• wandb: Whether Weights and Biases should be used for monitoring during training. Defaults to false.
• per_device_batch_size and dataloader_num_workers: The batch size and number of workers to use for training. Defaults to 8 and 4, respectively. Tweak these if you are running out of GPU memory.
• model.learning_rate, total_batch_size, max_steps, warmup_steps: Training parameters that you can tweak, although it shouldn't really be needed.

See all the finetuning options in the config/asr_finetuning.yaml file.

Evaluating an Automatic Speech Recognition (ASR) Model

You can use the evaluate_model script to evaluate an ASR model:

python src/scripts/evaluate_model.py [key=value]...

Here are some of the more important available keys:

• model_id (required): The Hugging Face model ID of the ASR model to evaluate.
• dataset: The ASR dataset to evaluate the model on. Can be any ASR dataset on the Hugging Face Hub. Note that subsets are separated with "::". For instance, to evaluate on the Danish Common Voice 17 dataset, you would use mozilla-foundation/common_voice_17_0::da. Defaults to alexandrainst/coral::read_aloud.
• eval_split_name: The dataset split to evaluate on. Defaults to test.
• text_column: The name of the column in the dataset that contains the text. Defaults to text.
• audio_column: The name of the column in the dataset that contains the audio. Defaults to audio.
• detailed: Only relevant if evaluating on the (default) CoRal test dataset. This will give a detailed evaluation across the different demographics in the dataset. If set to False it will only give the overall scores. Defaults to True.

See all the evaluation options in the config/evaluation.yaml file.

You can produce a comparison plot of different models evaluated on the CoRal test dataset with detailed=True by running the following script:

python src/scripts/create_comparison_plot.py \
  -f EVALUATION_FILE [-f EVALUATION_FILE ...] [--metric METRIC]

Here the EVALUATION_FILE arguments are the paths to the evaluation files produced by evaluate_model.py (they end in -coral-scores.csv). The METRIC argument is the metric to compare on, which can be one of wer and cer, for the word error rate and character error rate, respectively. The default is cer.

Troubleshooting

If you're on MacOS and get an error saying something along the lines of "fatal error: 'lzma.h' file not found" then try the following and rerun make install afterwards:

export CPPFLAGS="-I$(brew --prefix)/include"