Parakeet ASR Server

Local speech-to-text API using FluidAudio's Parakeet CoreML models. Achieves ~100-150x realtime transcription on Apple Silicon via the Neural Engine.

Requirements

• macOS 14.0+ (Sonoma)
• Apple Silicon (M1/M2/M3/M4)
• Xcode Command Line Tools (xcode-select --install)
• Python 3.11+ with uv

Setup

1. Build the transcribe binary

./scripts/build-transcribe.sh

This clones FluidAudio, adds the TranscribeCLI wrapper, and builds a release binary to bin/transcribe. The script is idempotent.

2. Download models (first run only)

The first transcription downloads ~1-2GB of CoreML models from HuggingFace. Do this once manually to avoid a slow first request:

./bin/transcribe data/test-2.wav

3. Install Python dependencies

uv sync

4. Run the server

uv run uvicorn server:app --host 0.0.0.0 --port 8765 --workers 2

Usage

Health check

curl http://localhost:8765/health

Transcribe audio

curl -X POST http://localhost:8765/transcribe -F "file=@data/test-1.wav"
curl -X POST http://localhost:8765/transcribe -F "file=@data/test-2.wav"

Response:

{
  "text": "Full transcription text",
  "segments": [
    { "start": 0.0, "end": 2.5, "text": "First sentence." },
    { "start": 2.5, "end": 5.0, "text": "Second sentence." }
  ],
  "confidence": 0.98,
  "rtfx": 155.0,
  "processing_time": 0.08
}

Redeploying

After pulling changes:

# If pyproject.toml changed
uv sync

# If scripts/TranscribeCLI.swift changed
./scripts/build-transcribe.sh

# Restart to pick up changes
./scripts/service.sh restart

For most changes (server.py), just restart is enough.

Running as a launchd service

./scripts/service.sh install   # Generate plist, install, and start
./scripts/service.sh status    # Check if running + health check
./scripts/service.sh stop
./scripts/service.sh start
./scripts/service.sh restart
./scripts/service.sh uninstall # Stop and remove

API

Endpoint	Method	Description
/health	GET	Health check
/transcribe	POST	Transcribe uploaded audio (wav, mp3, m4a, flac)
/docs	GET	Swagger UI

Architecture

HTTP Request → FastAPI (server.py) → subprocess → bin/transcribe → CoreML Neural Engine

The server shells out to the Swift binary for each transcription. The binary loads FluidAudio's Parakeet TDT v3 model and runs inference on the Apple Neural Engine.