---
language: mt
license: apache-2.0
library_name: transformers
tags:
  - whisper
  - automatic-speech-recognition
  - common-voice
  - mozilla-data-collective
metrics:
  - wer
base_model: openai/whisper-small
---

# openai/whisper-small fine-tuned on Common Voice 26.0 — Maltese

`openai/whisper-small` fine-tuned for Maltese (mt) automatic speech recognition
on Mozilla Common Voice Scripted Speech 26.0, trained on a laptop
(Apple M3 Pro, PyTorch MPS backend, fp32).

## Results

| Model | Raw WER | Normalized WER | Test clips |
|---|---|---|---|
| whisper-base fine-tuned | 105.7% | 94.1% | 300 |
| whisper-base zero-shot | 107.6% | 134.7% | 300 |
| whisper-small fine-tuned | 46.7% | 38.9% | 300 |
| whisper-small zero-shot | 110.1% | 100.6% | 300 |

Zero-shot Whisper barely functions on Maltese (≈1 h of it in the
original training mix); the WER delta above is what a few hours of
community-donated Common Voice speech and a laptop fine-tune buy. WER
computed on the official Common Voice test split (speaker-disjoint from
training data — note the training split draws on only 6
speakers, which limits generalization; see Limitations). "Normalized"
applies `BasicTextNormalizer` (lowercase, strip punctuation) to both
reference and hypothesis.

## Data provenance

- **Source:** Mozilla Common Voice Scripted Speech **26.0**,
  Maltese (mt), obtained via the
  [Mozilla Data Collective](https://datacollective.mozillafoundation.org/)
  (since October 2025 the exclusive distribution channel for Common Voice
  datasets — the former `mozilla-foundation/common_voice_*` datasets on this
  Hub are empty shells).
- **License:** the underlying speech data is **CC0**.
- **Used here:** 3000 training clips (~3.77 h,
  6 speakers), 200 dev clips, 300 test clips,
  official splits, clips 1–15 s.
- The prepared audio dataset is deliberately **not** re-uploaded to this Hub;
  get the data from the Mozilla Data Collective.

## Training

Standard Hugging Face `Seq2SeqTrainer` recipe
([fine-tune-whisper](https://huggingface.co/blog/fine-tune-whisper)):
batch 4 × grad-accum 4, lr 1e-05, warmup 50,
1000 steps, fp32 (MPS), gradient checkpointing, greedy decoding,
best-checkpoint-by-WER selection.

## Limitations

- Single low-resource language, ~3.77 h of scripted (read) speech
  from only **6 speakers** (the corpus's many one-clip
  contributors are concentrated in the held-out splits) — expect degradation
  on spontaneous/conversational audio and unseen voices.
- Small test set (300 clips); WER has meaningful variance.
- No punctuation/casing guarantees; laptop-scale training run, not a
  production model.

## Usage

```python
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="titaniumbones/whisper-small-mt-commonvoice")
print(asr("clip.mp3", generate_kwargs={"language": "maltese"}))
```