Whisper language benchmark table

Original FLEURS results

Values below are percentages from Table 13. Chinese, Japanese, Thai, Lao and Myanmar use character splitting in the paper’s normalizer and therefore report CER. Other rows report WER. The original experiment is not a measurement of quantized browser inference. The lowest number in a row can help you choose a sample test; it cannot promise transcript quality.

Use the table responsibly

Do not average CER and WER together as if they measured the same unit. Do not subtract an error rate from a hundred and call the result a guaranteed accuracy percentage. Insertions can make an error rate exceed a hundred. The denominator is the number of reference units, not all words the model emitted.

The paper used a specific collection of recordings and text normalization. Different accents, microphones, topics and turn-taking patterns can change performance. A table of read speech is not a direct estimate for a noisy meeting. These results are useful because the model sizes share a published evaluation setup. They are less useful as a forecast for an individual file.

Open a language page to see its chart, any available Common Voice result, and training-hour context. Then choose a representative passage containing names, pauses and background conditions similar to the material you actually need. Review the same passage with each candidate setting. Count corrections and assess whether the remaining mistakes would change meaning.

Keep the original recording as the reference. A draft can help locate passages, but an unreviewed transcript should not become the only record of a conversation. If a language has very high published error rates, manual transcription or a different system may be the more practical choice.