Audio scribe

AudioScribe

Speech-to-text with three swappable backends and one specialised endpoint for multi-track Audacity / Craig recordings. All endpoints accept a multipart file upload and return a Markdown transcript.

---

Base URL

Use whatever base URL the runtime gives you for AudioScribe; examples below use $BASE. Ask if it isn't configured.

---

Endpoints

All under /api/v1/transcribe. Each takes a file form field plus a few optional form fields. By default the response is the transcript file itself (Markdown). Set stream=true to get an SSE progress stream that ends with a download_url you can fetch from /api/v1/transcribe/download/{file_id}.

• POST /local: local faster-whisper on the host GPU. Best when the file is sensitive (no third-party API), GPU is available, or you want timestamps. Form fields: file, model (default Systran/faster-whisper-large-v3), language (auto if omitted), with_timestamps (default false), stream.
• POST /openai: OpenAI Whisper API. Best for short clips when you want the highest-quality general-purpose result and don't need timestamps. Server handles chunking for files >25MB. Form fields: file, model (default whisper-1), language, stream.
• POST /hf: Hugging Face Inference. Use when the user names a specific HF model or you want to avoid OpenAI. Form fields: file, model (default openai/whisper-large-v3), hf_provider (default hf-inference), stream.
• POST /audacity: multi-track .zip (Audacity .aup project, or a Craig Bot dump). Extracts each track, transcribes them with whichever backend you pick, and merges chronologically into [HH:MM:SS] [Speaker] … lines. Form fields: file (must be .zip), provider (local / openai / hf, default local), model, language, hf_provider, stream.

---

Picking a backend

Default to local when the host has a GPU, it's free, private, and supports timestamps. Switch to openai for short clips where quality matters and the recording is non-sensitive. Use hf only when the user pins a specific HF model. For multi-speaker or Discord/Craig recordings always use /audacity.

---

Examples

Quick transcription with the local model (response is the .md file), Bash:

curl -s -o transcript.md \
  -X POST $BASE/api/v1/transcribe/local \
  -F "[email protected]" \
  -F "with_timestamps=true"

PowerShell:

curl.exe -s -o transcript.md `
  -X POST $BASE/api/v1/transcribe/local `
  -F "[email protected]" `
  -F "with_timestamps=true"

OpenAI backend, English forced, Bash:

curl -s -o transcript.md \
  -X POST $BASE/api/v1/transcribe/openai \
  -F "[email protected]" \
  -F "language=en"

PowerShell:

curl.exe -s -o transcript.md `
  -X POST $BASE/api/v1/transcribe/openai `
  -F "[email protected]" `
  -F "language=en"

Audacity / Craig multi-track zip with chronological speaker merge, Bash:

curl -s -o transcript.md \
  -X POST $BASE/api/v1/transcribe/audacity \
  -F "[email protected]" \
  -F "provider=local"

PowerShell:

curl.exe -s -o transcript.md `
  -X POST $BASE/api/v1/transcribe/audacity `
  -F "[email protected]" `
  -F "provider=local"

Streaming progress (SSE) for a long file, Bash:

curl -N -X POST $BASE/api/v1/transcribe/local \
  -F "[email protected]" \
  -F "stream=true"

PowerShell:

curl.exe -N -X POST $BASE/api/v1/transcribe/local `
  -F "[email protected]" `
  -F "stream=true"

The stream emits {"type":"progress",…} events and ends with {"type":"complete","download_url":"/api/v1/transcribe/download/<id>"}. Fetch the URL from $BASE to get the Markdown.

Note for PowerShell: line continuation is the backtick `, not \. Use curl.exe so PowerShell doesn't route to its Invoke-WebRequest alias.

---

Tips

• Long files take real time: set the HTTP client timeout high (10+ minutes) for hour-long recordings, and prefer stream=true so the user sees progress instead of staring at a hung request.
• Pass language when you know it; auto-detection wastes the first chunk on identification.
• The Markdown response is plain text: feel free to post-process (summarise, extract action items, diarize further) once you have it.
• Servers' OPENAI_API_KEY / HF_TOKEN are configured server-side: don't ask the user for keys.