Parakeet TDT (Speech-to-Text)

Local transcription using NVIDIA Parakeet TDT 0.6B v3 with ONNX Runtime. Runs on CPU — no GPU required. ~30x faster than realtime.

Installation

# Clone the repo
git clone https://github.com/groxaxo/parakeet-tdt-0.6b-v3-fastapi-openai.git
cd parakeet-tdt-0.6b-v3-fastapi-openai

# Run with Docker (recommended)
docker compose up -d parakeet-cpu

# Or run directly with Python
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 5000

Default port is 5000. Set PARAKEET_URL to override (e.g., http://localhost:5092).

API Endpoint

OpenAI-compatible API at $PARAKEET_URL (default: http://localhost:5000).

Quick Start

# Transcribe audio file (plain text)
curl -X POST $PARAKEET_URL/v1/audio/transcriptions \
  -F "file=@/path/to/audio.mp3" \
  -F "response_format=text"

# Get timestamps and segments
curl -X POST $PARAKEET_URL/v1/audio/transcriptions \
  -F "file=@/path/to/audio.mp3" \
  -F "response_format=verbose_json"

# Generate subtitles (SRT)
curl -X POST $PARAKEET_URL/v1/audio/transcriptions \
  -F "file=@/path/to/audio.mp3" \
  -F "response_format=srt"

Python / OpenAI SDK

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.getenv("PARAKEET_URL", "http://localhost:5000") + "/v1",
    api_key="not-needed"
)

with open("audio.mp3", "rb") as f:
    transcript = client.audio.transcriptions.create(
        model="parakeet-tdt-0.6b-v3",
        file=f,
        response_format="text"
    )
print(transcript)

Response Formats

Format	Output
text	Plain text
json	{"text": "..."}
verbose_json	Segments with timestamps and words
srt	SRT subtitles
vtt	WebVTT subtitles

Supported Languages (25)

English, Spanish, French, German, Italian, Portuguese, Polish, Russian, Ukrainian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Czech, Romanian, Hungarian, Bulgarian, Slovak, Croatian, Lithuanian, Latvian, Estonian, Slovenian

Language is auto-detected — no configuration needed.

Web Interface

Open $PARAKEET_URL in a browser for drag-and-drop transcription UI.

Docker Management

# Check status
docker ps --filter "name=parakeet"

# View logs
docker logs -f <container-name>

# Restart
docker compose restart

# Stop
docker compose down

Why Parakeet over Whisper?

• Speed: ~30x faster than realtime on CPU
• Accuracy: Comparable to Whisper large-v3
• Privacy: Runs 100% locally, no cloud calls
• Compatibility: Drop-in replacement for OpenAI's transcription API