Whisper-large-et

This is a Whisper-large-v2 model openai/whisper-large-v2 finetuned on around 1200 hours of diverse Estonian data.

Model description

This is a general-purpose Estonian ASR model trained in the Lab of Language Technology at TalTech.

Intended uses & limitations

This model is intended for general-purpose speech recognition, such as broadcast conversations, interviews, talks, etc.

How to use

Recommended: use faster-whisper.

For example:

Convert the HF model to CT2 format:

ct2-transformers-converter --model TalTechNLP/whisper-large-et --output_dir whisper-large-et.ct2 --copy_files tokenizer.json --quantization float16
Decode:

whisper-ctranslate2 --model_directory whisper-large-et.ct2 --task transcribe --language et --beam_size 5 some_file.mp3

Limitations and bias

Since this model was trained on mostly broadcast speech and texts from the web, it might have problems correctly decoding the following:

Speech containing technical and other domain-specific terms
Children's speech
Non-native speech
Speech recorded under very noisy conditions or with a microphone far from the speaker
Very spontaneous and overlapping speech

Training data

Acoustic training data:

Type	Amount (h)
Broadcast speech	991
Spontaneous speech	53
Elderly speech corpus	53
Talks, lectures	49
Parliament speeches	31
Total	1161

Training procedure

Finetuned using Espnet, and then comverted to transformers format using this script. Finetuning procedure is similar to this model. Finetuning was done for 3 epochs, with model averaging at the end of training.

Update: 2023-10-03 version of the model is trained on long segments (like the original Whisper model) and is therefore especially well suited to be used e.g. with faster-whisper to transcribe long speech recordings "end-to-end" (i.e., without any prior segmentation).

Evaluation results

WER

WER results below are obtained using greedy decoding (i.e., beam size 1).

Dataset	WER
Common Voice 8.0	11.3
Common Voice 11.0	12.0