Stock Whisper Base reaches a 62.50% word error rate (WER) on held-out Greek speech. Whisper Base el-en reaches 28.65% with the same 72.6 million parameters. English became less accurate in exchange, and this article gives both numbers.
Why a small model
We wanted Greek dictation on a laptop processor, so the model runs through whisper.cpp with q8_0 weights. Large Whisper models are more accurate on Greek, and they need a bigger machine. Whisper Base el-en is for the cases where the audio has to stay on a local CPU.
What we trained
We fine-tuned the full model on 179.2 hours of Greek and 90.0 hours of English. Training draws were 70% Greek and 30% English. Greek references keep their orthography as written: tonos, dialytika, final sigma, capitalization, and punctuation. A missing tonos counts as an error in every Greek score below.
The run took 8,000 updates of 64 clips and 65 minutes on one RTX 4090. We selected the checkpoint at step 6,000 on FLEURS validation and scored the test split once.
Greek results
Greek word error rate
| Model | Parameters | WER |
|---|---|---|
| Whisper Base el-en | 72.6 M | 28.65% |
| Whisper Base | 72.6 M | 62.50% |
| Whisper Small | 241.7 M | 35.11% |
| Whisper Medium | 763.9 M | 20.77% |
| Whisper large-v3 | 1,543.5 M | 11.45% |
Lexical WER on the held-out FLEURS Greek test split. Lower is better. Accents count.
The fine-tune beats stock Whisper Small, which has 3.3 times as many parameters. Medium and larger stock models stay ahead.
English from Greek speakers
Stock Whisper often hears Greek when a Greek speaker reads English. On read English by 11 Greek and Cypriot speakers, with detection restricted to the two languages, every stock model wrote Greek for one to three speakers. Whisper Base el-en decoded all 11 as English and reached 11.86% WER. With 11 speakers, no stock model is statistically separable from it.
What got worse
Native English lost accuracy. Mean WER over nine public Open ASR Leaderboard sets rose from 12.09% to 15.58%, and every set got worse. With English forced, stock Whisper Base is the better choice. The model also deletes the final word of 7.4% of Greek validation clips, and silence can produce short invented phrases without voice activity detection.
Use the model
The weights are on Hugging Face under Apache-2.0, in Transformers and whisper.cpp formats, with the technical report. The model page lists the specifications and limits. Every result comes from one training run with one seed, so test your own audio before you deploy it.