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

# Audio formats and preprocessing

> Supported audio formats across all Modulate model endpoints, with guidance on format selection, conversion, and the special requirements of the streaming Deepfake Detection endpoint.

Reference for supported audio formats across all Modulate model endpoints, with guidance on format selection, conversion, and the special requirements of the streaming Deepfake Detection endpoint.

## Format support by endpoint

Every endpoint advertises its accepted `audio_format` values in its spec. The spec is the authoritative list — check the API reference for each endpoint to confirm exact format support, since formats are added over time. Broadly:

* **HTTP batch endpoints (Multilingual Transcription, English Fast Transcription, Multilingual Fast Transcription, Deepfake Detection, PII/PHI Redaction)** accept common container formats; the file extension determines decoding. See each spec for the current list.
* **Multilingual Transcription streaming and PII/PHI Redaction streaming** accept self-describing container formats (auto-detected from headers) and raw PCM / mu-law / A-law (which require `audio_format`, `sample_rate`, and `num_channels`).
* **Deepfake Detection streaming** has the broadest container support plus raw PCM, with required `audio_format` and conditional `sample_rate` / `num_channels` for raw formats.

<Warning>
  Recommended maximum file size for all HTTP batch endpoints: **100 MB**.
</Warning>

## Choosing a format

**Opus** is the recommended format for streaming use cases. It provides excellent audio quality at low bitrates, reducing bandwidth consumption while preserving the acoustic detail the models need.

For batch endpoints, any supported container format works — use whatever format your audio pipeline already produces.

## Deepfake Detection streaming: raw PCM vs container formats

The Deepfake Detection streaming endpoint accepts two categories of audio format, declared via the `audio_format` query parameter.

### Container formats

Container formats include metadata — sample rate, channel count, codec — within the stream itself. When using a container format, `sample_rate` and `num_channels` are not required. See the Deepfake Detection streaming spec's `audio_format` enum for the full list of accepted container values.

```text theme={null}
wss://...?api_key=YOUR_API_KEY&audio_format=webm
```

### Raw PCM formats

Raw formats are headerless audio samples. The server cannot infer sample rate or channel count from the data itself, so `sample_rate` and `num_channels` are **required** query parameters when using any raw format.

```text theme={null}
wss://...?api_key=YOUR_API_KEY&audio_format=s16le&sample_rate=16000&num_channels=1
```

Supported raw formats:

`s8`, `s16le`, `s16be`, `s24le`, `s24be`, `s32le`, `s32be`, `u8`, `u16le`, `u16be`, `u24le`, `u24be`, `u32le`, `u32be`, `f32le`, `f32be`, `f64le`, `f64be`, `mulaw`, `alaw`

#### Common raw format configurations

| Use case                       | `audio_format` | `sample_rate` | `num_channels` |
| ------------------------------ | -------------- | ------------- | -------------- |
| Default / native app           | `s16le`        | `16000`       | `1`            |
| Web Audio API (`AudioWorklet`) | `f32le`        | `48000`       | `1`            |
| Native stereo capture          | `s16le`        | `48000`       | `2`            |
| Telephony (mu-law)             | `mulaw`        | `8000`        | `1`            |
| Telephony (A-law)              | `alaw`         | `8000`        | `1`            |

### The `s16le` passthrough optimization

When the input is `s16le` at 16 kHz mono, no format conversion is performed before analysis. This is the most efficient configuration for the Deepfake Detection streaming endpoint.

All other formats are decoded and resampled to 16 kHz mono before analysis. There is no functional difference in output, but the passthrough path avoids the conversion overhead.

<Tip>
  If you control the audio capture pipeline and are integrating with the Deepfake Detection streaming endpoint, capture in `s16le` at 16 kHz mono to take advantage of zero-cost passthrough.
</Tip>

### Supported `sample_rate` values

`8000`, `11025`, `16000`, `22050`, `32000`, `44100`, `48000`, `96000`

### `num_channels` range

1–8. Multi-channel audio is downmixed to mono before analysis.

## Error handling for format problems

| Endpoint                                | Status / code     | Cause                                                                    |
| --------------------------------------- | ----------------- | ------------------------------------------------------------------------ |
| English Fast Transcription (batch)      | `400`             | Unsupported file extension, empty file, or decode error                  |
| Multilingual Transcription (batch)      | `400`             | Unsupported format or empty file                                         |
| Multilingual Fast Transcription (batch) | `400`             | Unsupported format, empty file, or unprocessable audio                   |
| Deepfake Detection (batch)              | `400`             | Empty file or unsupported format                                         |
| Deepfake Detection (batch)              | `422`             | Audio shorter than 0.5 seconds                                           |
| Deepfake Detection (streaming)          | Close code `1003` | Invalid `audio_format`, `sample_rate`, or `num_channels` query parameter |
| Deepfake Detection (streaming)          | Close code `4002` | Audio could not be decoded or does not match the declared format         |

<Warning>
  The Deepfake Detection streaming endpoint validates format parameters at connection time (close code `1003`) and again when it attempts to decode audio (close code `4002`). This covers both undecodable audio (e.g. a truncated or corrupt stream) and format mismatches (e.g. declaring `audio_format=s16le` but sending WebM data).
</Warning>

## Related

* [Code examples by language](/guides/code-examples) — includes Opus conversion patterns
* [Troubleshooting](/guides/troubleshooting)