Overview
Podonos evaluates speech with native speakers in the target locale. Use the language code below as thelan argument when creating an evaluator (see create_evaluator()).
Supported languages
Here is a list of supported languages and their median turnaround time.| Flag | Code | Language | Turnaround |
|---|---|---|---|
| 🇺🇸 | en-us | English (United States) | 6h 10m |
| 🇬🇧 | en-gb | English (United Kingdom) | 5h 49m |
| 🇦🇺 | en-au | English (Australia) | 4h 34m |
| 🇨🇦 | en-ca | English (Canada) | 12h |
| 🇮🇳 | en-in | English (India) | 4h 15m |
| 🇰🇷 | ko-kr | Korean (Korea) | 5h 7m |
| 🇨🇳 | zh-cn | Mandarin (China) | 7h 31m |
| 🇪🇸 | es-es | Spanish (Spain) | 4h 33m |
| 🇲🇽 | es-mx | Spanish (Mexico) | 5h 46m |
| 🇫🇷 | fr-fr | French (France) | 3h 46m |
| 🇨🇦 | fr-ca | French (Canada) | 12h |
| 🇩🇪 | de-de | German (Germany) | 5h 40m |
| 🇯🇵 | ja-jp | Japanese (Japan) | 14h 27m |
| 🇮🇹 | it-it | Italian (Italy) | 2h 17m |
| 🇵🇱 | pl-pl | Polish (Poland) | 12h |
| 🇵🇹 | pt-pt | Portuguese (Portugal) | 3h 3m |
| 🇧🇷 | pt-br | Portuguese (Brazil) | 5h 45m |
| 🇱🇰 | si-lk | Sinhala (Sri Lanka) | 48h |
| 🇮🇳 | ta-in | Tamil (India) | 48h |
| 🇮🇳 | kn-in | Kannada (India) | 48h |
| 🇮🇳 | ml-in | Malayalam (India) | 48h |
| 🇮🇳 | hi-in | Hindi (India) | 48h |
| 🇪🇬 | ar-eg | Arabic (Egypt) | 48h |
| 🇦🇪 | ar-ae | Arabic (UAE) | 48h |
| 🇸🇦 | ar-sa | Arabic (Saudi Arabia) | 48h |
| 🎧 | audio | General audio file | 12h |
About these turnaround times
The figures above are measured from real evaluations completed within the last 3 months, not theoretical targets. Across all locales, the overall median return time was 5h 56m and the overall average was 23h 28m. The Turnaround column reports the median for each locale. A few things to keep in mind:- Some languages take longer. Sinhala, Tamil, Kannada, Malayalam, Hindi, and Arabic locales have smaller evaluator pools and lower throughput, so individual runs can extend beyond the 48-hour figure shown. We do not throttle them — we wait for qualified native speakers rather than relaxing the bar.
- Larger evaluations take proportionally longer. Figures assume a standard
num_evaland moderate sample count. Very large jobs extend the window. - Time of day and timezone matter. A job launched as the target region wakes up returns faster than one launched at local midnight.