> ## Documentation Index
> Fetch the complete documentation index at: https://docs.audivo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Languages

> Which languages are supported, what the support tiers mean, and when to pass one explicitly.

`GET /v1/languages` returns the supported-language table as a list of rows:

```json theme={null}
{
  "language": "en",
  "engine": "whisper-large-v3-turbo",
  "support_tier": "benchmarked",
  "max_timing_precision": "word"
}
```

One hundred languages are listed, one per row, with the same engine on every
row: the product runs one pinned engine, and automatic language detection
never selects another. Word timing is what that engine returns, so
`max_timing_precision` is `word` throughout.

## What the tiers mean

The tiers say what Audivo has **measured**, not what the engine claims:

| tier          | meaning                                                                                                                                                                                            |
| ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `benchmarked` | Audivo's own provider benchmark scored the engine on real episodes in this language. English, today.                                                                                               |
| `supported`   | The engine documents the language. Audivo has run it, but has not measured word accuracy on real episodes.                                                                                         |
| `best_effort` | Reserved for a documented language Audivo has evidence of degraded output on. No language is listed at this tier yet; the tier exists so evidence has somewhere to land without a contract change. |

## When to pass `language`

Omit it and the engine detects the language from the audio, records it on the
transcript, and flags its own doubts as warnings (`language_mismatch`,
`low_language_confidence`, `language_unresolved` — see
[Output formats](/formats)). Pass a BCP-47 tag on `POST /v1/transcripts`
when you already know the language, and always for audio that mixes
languages or opens with music, where detection is least reliable. The tag is
checked for shape only; one the engine does not document is not refused
today, and its output is unmeasured.