Measured 2026-09-17 on FLEURS
Pikka Speech vs OpenAI gpt-transcribe: which is more accurate?
Across 84 languages, Pikka Speech records a median character error rate of 2.9% against OpenAI gpt-transcribe’s 4.5% on the same audio — lower is better. Pikka Speech wins or ties 65 of 84.
Per language
Character error rate, language by language
Lower is better. Difference is Pikka Speech minus OpenAI gpt-transcribe; a negative value means Pikka Speech is ahead.
| Language | Pikka Speech | OpenAI gpt-transcribe | Difference |
|---|---|---|---|
| Afrikaans | 2.8% | 3.0% | -0.2 pts |
| Amharic | 3.9% | 14.9% | -11.0 pts |
| Arabic | 4.1% | 5.4% | -1.3 pts |
| Armenian | 12.3% | 14.2% | -1.9 pts |
| Assamese | 8.5% | 12.3% | -3.8 pts |
| Asturian | 7.6% | 9.4% | -1.8 pts |
| Azerbaijani | 0.9% | 2.5% | -1.6 pts |
| Belarusian | 2.2% | 2.8% | -0.6 pts |
| Bengali | 0.9% | 3.2% | -2.3 pts |
| Bulgarian | 0.4% | 0.4% | 0.0 pts |
| Burmese | 9.9% | 41.6% | -31.7 pts |
| Cantonese | 39.6% | 4.8% | +34.8 pts |
| Catalan | 2.2% | 3.7% | -1.5 pts |
| Cebuano | 9.8% | 8.3% | +1.5 pts |
| Chinese | 4.9% | 6.8% | -1.9 pts |
| Croatian | 0.4% | 27.8% | -27.4 pts |
| Czech | 6.0% | 4.2% | +1.8 pts |
| Danish | 1.8% | 3.3% | -1.5 pts |
| Dutch | 1.8% | 0.4% | +1.4 pts |
| English | 3.1% | 1.5% | +1.6 pts |
| Estonian | 0.5% | 1.9% | -1.4 pts |
| Finnish | 0.3% | 0.6% | -0.3 pts |
| French | 3.9% | 2.5% | +1.4 pts |
| Galician | 1.1% | 1.5% | -0.4 pts |
| Georgian | 4.1% | 6.8% | -2.7 pts |
| German | 5.1% | 0.4% | +4.7 pts |
| Greek | 2.4% | 2.1% | +0.3 pts |
| Gujarati | 3.4% | 8.3% | -4.9 pts |
| Hausa | 5.9% | 17.4% | -11.5 pts |
| Hebrew | 6.7% | 6.1% | +0.6 pts |
| Hindi | 1.7% | 3.6% | -1.9 pts |
| Hungarian | 4.0% | 2.3% | +1.7 pts |
| Icelandic | 10.9% | 12.2% | -1.3 pts |
| Indonesian | 1.1% | 2.6% | -1.5 pts |
| Irish | 22.8% | 32.4% | -9.6 pts |
| Italian | 0.3% | 0.6% | -0.3 pts |
| Japanese | 1.8% | 3.6% | -1.8 pts |
| Javanese | 3.1% | 11.4% | -8.3 pts |
| Kannada | 1.6% | 4.6% | -3.0 pts |
| Kazakh | 0.9% | 4.5% | -3.6 pts |
| Khmer | 9.6% | 12.3% | -2.7 pts |
| Korean | 2.2% | 2.9% | -0.7 pts |
| Kyrgyz | 5.6% | 5.7% | -0.1 pts |
| Lao | 17.6% | 62.0% | -44.4 pts |
| Latvian | 2.4% | 4.0% | -1.6 pts |
| Lithuanian | 2.1% | 6.5% | -4.4 pts |
| Luxembourgish | 35.5% | 16.9% | +18.6 pts |
| Macedonian | 1.0% | 1.3% | -0.3 pts |
| Malay | 2.2% | 1.3% | +0.9 pts |
| Malayalam | 1.2% | 4.3% | -3.1 pts |
| Maltese | 12.0% | 6.7% | +5.3 pts |
| Marathi | 3.7% | 7.0% | -3.3 pts |
| Mongolian | 5.2% | 12.5% | -7.3 pts |
| Nepali | 0.0% | 7.0% | -7.0 pts |
| Northern Sotho | 13.2% | 34.7% | -21.5 pts |
| Norwegian | 0.7% | 1.0% | -0.3 pts |
| Odia | 7.8% | 10.6% | -2.8 pts |
| Persian | 1.9% | 2.3% | -0.4 pts |
| Polish | 1.6% | 1.0% | +0.6 pts |
| Portuguese | 1.0% | 2.8% | -1.8 pts |
| Punjabi | 4.6% | 23.9% | -19.3 pts |
| Romanian | 2.9% | 0.8% | +2.1 pts |
| Russian | 0.2% | 0.2% | 0.0 pts |
| Serbian | 15.1% | 88.1% | -73.0 pts |
| Sindhi | 4.2% | 98.5% | -94.3 pts |
| Slovak | 1.2% | 1.8% | -0.6 pts |
| Slovenian | 2.0% | 3.5% | -1.5 pts |
| Spanish | 1.2% | 0.4% | +0.8 pts |
| Swahili | 1.9% | 5.3% | -3.4 pts |
| Swedish | 4.5% | 3.8% | +0.7 pts |
| Tagalog | 2.3% | 4.9% | -2.6 pts |
| Tajik | 2.9% | 58.3% | -55.4 pts |
| Tamil | 16.8% | 18.6% | -1.8 pts |
| Telugu | 1.4% | 3.4% | -2.0 pts |
| Thai | 4.0% | 4.9% | -0.9 pts |
| Turkish | 0.7% | 1.0% | -0.3 pts |
| Ukrainian | 1.2% | 3.3% | -2.1 pts |
| Urdu | 4.7% | 90.5% | -85.8 pts |
| Uzbek | 7.7% | 9.2% | -1.5 pts |
| Vietnamese | 1.1% | 3.6% | -2.5 pts |
| Welsh | 5.7% | 13.2% | -7.5 pts |
| Xhosa | 49.0% | 33.7% | +15.3 pts |
| he | 6.7% | 6.1% | +0.6 pts |
| pt-BR | 1.0% | 2.8% | -1.8 pts |
Both sides
Where OpenAI gpt-transcribe leads
19 of 84 languages, listed in full — the benchmark publishes its losses, not only its wins.
- Cantonese: OpenAI gpt-transcribe 4.8% vs Pikka Speech 39.6%
- Luxembourgish: OpenAI gpt-transcribe 16.9% vs Pikka Speech 35.5%
- Xhosa: OpenAI gpt-transcribe 33.7% vs Pikka Speech 49.0%
- Maltese: OpenAI gpt-transcribe 6.7% vs Pikka Speech 12.0%
- German: OpenAI gpt-transcribe 0.4% vs Pikka Speech 5.1%
- Romanian: OpenAI gpt-transcribe 0.8% vs Pikka Speech 2.9%
- Czech: OpenAI gpt-transcribe 4.2% vs Pikka Speech 6.0%
- Hungarian: OpenAI gpt-transcribe 2.3% vs Pikka Speech 4.0%
- English: OpenAI gpt-transcribe 1.5% vs Pikka Speech 3.1%
- Cebuano: OpenAI gpt-transcribe 8.3% vs Pikka Speech 9.8%
- Dutch: OpenAI gpt-transcribe 0.4% vs Pikka Speech 1.8%
- French: OpenAI gpt-transcribe 2.5% vs Pikka Speech 3.9%
- Malay: OpenAI gpt-transcribe 1.3% vs Pikka Speech 2.2%
- Spanish: OpenAI gpt-transcribe 0.4% vs Pikka Speech 1.2%
- Swedish: OpenAI gpt-transcribe 3.8% vs Pikka Speech 4.5%
- Hebrew: OpenAI gpt-transcribe 6.1% vs Pikka Speech 6.7%
- Polish: OpenAI gpt-transcribe 1.0% vs Pikka Speech 1.6%
- he: OpenAI gpt-transcribe 6.1% vs Pikka Speech 6.7%
- Greek: OpenAI gpt-transcribe 2.1% vs Pikka Speech 2.4%
FAQ
OpenAI gpt-transcribe vs Pikka Speech
How is Pikka Speech compared against OpenAI gpt-transcribe?
Both engines transcribed the same FLEURS audio — 10 utterances per language — through their own production interfaces, and results are scored as character error rate (CER) after normalising case and punctuation. Lower is better. Full per-language detail is on the speech accuracy page.
Where does OpenAI gpt-transcribe win?
OpenAI gpt-transcribe records a lower error rate on 19 of the 84 languages compared — Cantonese (4.8% vs 39.6%), Luxembourgish (16.9% vs 35.5%), Xhosa (33.7% vs 49.0%), Maltese (6.7% vs 12.0%), German (0.4% vs 5.1%). Those rows are published in full on this page.
What does a difference in character error rate mean in practice?
CER is the share of characters an engine gets wrong: 3% means roughly three characters in every hundred differ from what was actually said. In live captions and interpretation, those errors concentrate in names, brands and technical terms — which is why the measured difference matters most on event audio.
Keep exploring
The full benchmark across all engines: speech recognition accuracy. Which languages we cover: language coverage. Or start a free test room.
Hear the difference on your own event
Create a free test room, speak a few sentences, and compare the live captions against OpenAI gpt-transcribe on your own audio.