plaud-embedded-transcription-api-skill

The saved excerpt is a snapshot from review. The external source remains the complete and most current version.

---
name: plaud-embedded-transcription-api-skill
description: Skill for users to implement Plaud Embedded's Transcription API. Use this skill when the user wants to transcribe audio files from their Plaud device or mobile app.
---

# Plaud Embedded Transcription API Skill
This skill provides context and instructions on how to upload audio files to Plaud for transcription.

## When To Use This Skill
Use this skill when a user wants to transcribe audio files using Plaud's Transcription Pipeline (the pipeline performs language detection, 
noise reduction, ASR, Speech-to-text, etc.)

### Prerequisites
- [ ] User has audio files on either: 
    1. Their mobile app synced from a Plaud Device
    2. An audio file available for download on a public URL (i.e. S3 download URL)

## How Plaud's Transcription API works

The **Transcription API** is an AI API with two main endpoints.

1. Triggers an async transcription job given a `file_url` to download an audio file
2. Get transcription task status for polling; On task finish, the conversation transcription will be available

## How to Start Transcribing

The Transcription API itself requires a public download URL to download the audio file. 

**If the developer does not have a way to upload files and generate public download URLs already, use the File Upload API**.

Else, skip the File Upload API and use the Transcription API directly.

### File Upload API \[ONLY FOR USERS WHO DO NOT HAVE A DOWNLOAD URL TO ACCESS THEIR AUDIO FILE\]

The File Upload API is an API to use Plaud's cloud storage to upload audio files. 
This can be performed either from your mobile app or from your backend.

Read the [File Upload Overview](https://docs.plaud.ai/plaud-embedded/file-api-overview.md) for the available File Upload API endpoints and data flow.

It is a **3-step multipart upload**:

1. `POST /generate-presigned-urls` with `filesize` and `filetype` → returns `FileId`, `UploadId`, `ChunkSize`, and a `Parts` array of presigned S3 URLs
2. `PUT` up to `ChunkSize` of raw bytes to each `PresignedUrl` (no auth — these are presigned). **Keep the `ETag` response header from every `PUT`** — the next step needs them
3. `POST /complete-upload` with `file_id`, `upload_id`, the `part_list` of `PartNumber`/`ETag` pairs, `filetype`, and `file_md5` → returns the `DownloadUrl`

**IMPORTANT**: The returned `DownloadUrl` is valid for **24 hours**. Pass it as `file_url` to the Transcription API.

#### API Reference (The Upload Step is not included as it's directly to S3)
* [Generating presigned upload URLs](https://docs.plaud.ai/api-reference/file-upload-api/generate-presigned-upload-urls.md) 
* [Completing the upload](https://docs.plaud.ai/api-reference/file-upload-api/complete-multipart-upload.md) 

### Transcription API
After an audio file has been uploaded to a public download API (either via the File Upload API or through a user's unique cloud storage), the user can use the Transcription API on the file URL.

The [Transcription API Overview](https://docs.plaud.ai/plaud-embedded/transcription-api-overview.md) goes through how Plaud's transcription flow works.

**IMPORTANT**: The Transcription API authenticates with your `X-Client-Id` and `X-Client-Api-Key` headers (the `api_key` is NOT your `client_secret` — grab it from the developer portal under App Settings > API Keys). 

Supported audio formats for `file_url` are **M4A, MP3, and WAV**. Recordings **exceeding 5 hours** should be broken into chunks and transcribed in parts.

`POST` accepts an optional `params` object to tune the pipeline:

| Param | Default | Purpose |
| ---- | ---- | ---- |
| `transcribe.language` | `auto` | BCP-47 code (`en-US`, `zh-CN`) or `auto` |
| `transcribe.detection_level` | `segment` | Language identification level (`segment` or `chapter`) |
| `vad.decode_silence` | `false` | Whether to decode silent regions |
| `diarization.enabled` | `false` | Identify and label speakers |
| `diarization.return_embedding` | `false` | Return speaker embed