pallr-whisper-tiny-is: small Icelandic speech recognition

Whisper-tiny (39M parameters) fine-tuned for Icelandic, small enough for phones. Writes cased, punctuated text.

Author: Páll Rúnarsson, HUMI lab, Reykjavík University.

Training data: Samrómur Milljón, Samrómur 21.05 and Spjallrómur.

Builds on the Samrómur-trained models from the article Scaling Smaller ASR Models Against Multilingual ASR Giants for Low-Resource Languages: A Comprehensive Evaluation for Icelandic (Language and Voice Lab / HUMI lab, Reykjavík University; model collection).

Results

Word error rate (WER, %, lower is better), lowercase with punctuation removed before scoring:

Test set	WER
Samrómur 21.05 test	19.2
Fjölmiðlatal 22.02 (RÚV radio)	46.9
Samrómur Milljón test	5.5

The Samrómur Milljón test set contains sentences that also occur in the Milljón training prompts, so its score is in-domain.

Compared with the large Icelandic models

Model	Parameters	Weights	Samrómur 21.05	Fjölmiðlatal
this model (pallr-whisper-tiny-is)	39M	~150 MB	19.2	46.9
whisper-large-icelandic-62640-steps-967h	1,550M	~6.2 GB	6.9	38.1
wav2vec2-large-xlsr-53-icelandic-ep30-967h	315M	~1.3 GB	7.0	45.4

The large models are more accurate on read speech, but they are hardware heavy. This model is a fraction of their size, so it loads in seconds and runs on an ordinary phone or laptop CPU, which makes offline transcription on the device possible.

How to use it in Python

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="palli23/pallr-whisper-tiny-is")
out = asr(
    "recording.wav",                      # any audio file; it is resampled to 16 kHz
    chunk_length_s=30,                    # needed for recordings longer than 30 seconds
    generate_kwargs={"language": "icelandic", "task": "transcribe",
                     "repetition_penalty": 1.2, "no_repeat_ngram_size": 3},
)
print(out["text"])

Decoding with repetition_penalty=1.2 and no_repeat_ngram_size=3 is recommended; without them Whisper can repeat itself on silence or noise.

You can also try it in your browser, with nothing uploaded, at ormal.is.

Licence and funding

CC BY-SA 4.0. Developed with sponsorship from Glóð and co-funded by the European Union.