parakeet-stt

# 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

```bash
# 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

```bash
# 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

```python
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, Croa...