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

# Gemini 3.5 Transcribe

> Google's dedicated file speech-to-text model. Pin `gemini-3.5-transcribe` on POST /v1/audio/transcriptions when you want Google file STT; the `transcribe` alias still points at Whisper.

## Overview

Gemini 3.5 Transcribe is a speech-to-text model for recorded audio, not a chat model and not a live streaming session. On Kyma it is served through the same OpenAI-compatible transcription endpoint as the other STT SKUs: multipart `file` or JSON `audio_url`, response `{ text, language?, duration? }` with `json` or `text` — it does not return per-segment timestamps.

It accepts up to about an hour of audio per request. Billing is per minute of audio (one-minute minimum, rounded up), not per token. The `transcribe` alias is unchanged and still resolves to whisper-v3-turbo; this SKU is opt-in by id.

This is file transcription only. It does not replace audio-scene understanding (`audio-understand`) and it is not the live streaming companion.

## Specs

| Field             | Value                                                |
| ----------------- | ---------------------------------------------------- |
| Model ID          | `gemini-3.5-transcribe`                              |
| Best for          | File transcription, multilingual dictation, captions |
| Context window    | 3,600                                                |
| Input modalities  | Audio                                                |
| Output modalities | Text                                                 |
| Speed             | Fast                                                 |
| Cost band         | Balanced                                             |
| Release stage     | Stable                                               |

## Pricing

|                     |   Price   |
| ------------------- | :-------: |
| Per minute of audio | \$0.00675 |

Billed per minute of audio. Every request reports its exact cost in `usage.cost`; the full media price list is on [Pricing](/pricing).

## Use this when

* **Meeting and call transcripts** — Turn recorded meetings and call logs into plain text for search and notes.
* **Multilingual dictation** — Transcribe speech across many languages without swapping SKUs.
* **Captions from a file** — Produce a transcript from an uploaded clip, then format it client-side.
* **Voice-agent offline pass** — Batch-transcribe recorded user audio before a later language-model step.

## Not ideal for

Do not use this for live bidirectional speech or for audio-scene questions (tone, music, SFX). Those are different endpoints. It returns a plain transcript with no per-segment timestamps: `verbose_json`, `srt` and `vtt` are refused on this SKU (`400 segments_unsupported`) — use whisper-v3-turbo for timestamps.

## Pick something else when

* You want the default cheap and fast alias: use [`whisper-v3-turbo`](/models/whisper-v3-turbo).
* You need premium conversational accuracy and opted into the quality alias: use [`gpt-4o-mini-transcribe-2025-12-15`](/models/gpt-4o-mini-transcribe-2025-12-15).

## Example

```python theme={null}
from openai import OpenAI

client = OpenAI(base_url="https://kymaapi.com/v1", api_key="kyma-...")

response = client.audio.transcriptions.create(
    model="gemini-3.5-transcribe",
    file=open("audio.wav", "rb"),
)
```

## Agent query example

Ask the API which models fit, instead of hardcoding an id:

```bash theme={null}
curl "https://kymaapi.com/v1/models?recommended_for=transcribe-audio&quality_tier=frontier-open"
```

## FAQ

**Does the transcribe alias point here?**

No. `transcribe` still resolves to whisper-v3-turbo. Pin `gemini-3.5-transcribe` when you want this model.

**How is this billed?**

Per minute of audio, rounded up, with a one-minute minimum. The generated Pricing table on this page is the live rate.

**Is this a chat completions model?**

No. Send audio to POST /v1/audio/transcriptions. A chat request is the wrong surface.