Parakeet-API: High-Performance Local Speech Transcription Service

What is this?

parakeet-api is a local speech transcription service based on the NVIDIA Parakeet-tdt-0.6b model. Key features:

• Provides an OpenAI API-compatible interface for direct integration with pyVideoTrans
• Built-in clean web user interface — open a browser to use it
• Transcription output in standard SRT subtitle format
• Runs entirely locally — no internet required, no payments, no privacy concerns
• Compatible with pyVideoTrans v3.72+

Open-source repository: https://github.com/jianchang512/parakeet-api

Use Cases

• Users who need offline speech recognition
• Users with high requirements for recognition speed and accuracy
• Users who prefer not to rely on cloud APIs and prioritize data privacy
• Users with an NVIDIA GPU (GPU acceleration provides a dramatic speed boost)

Prerequisites

Requirement	Details
NVIDIA GPU (recommended)	CUDA-compatible GPU significantly improves speed; CPU mode is available but slower
FFmpeg	For audio/video format preprocessing
Python 3.10	Required for source code deployment
pyVideoTrans v3.72+	Required version of the translation tool

Quick Start: Download the Bundle (Recommended for Beginners)

The bundle does not require installing Python or dependencies — just extract and run:

Download link 1 (Baidu Netdisk): Download from Baidu Netdisk

Download link 2 (HuggingFace): Download from HuggingFace.co

Steps

1. Download the archive and extract it to any folder (path must not contain Chinese characters or spaces)
2. Double-click the 启动.bat (Start) file
3. Wait for the command line window to show a startup success message — the browser will open automatically

Startup success screen:

Seeing the interface above means the service started successfully. You can use it directly in the browser or integrate it with pyVideoTrans.

Using in pyVideoTrans

Parakeet-API integrates seamlessly with pyVideoTrans (v3.72 and above).

Configuration Steps

1. Ensure parakeet-api is running locally (keep the command line window open)
2. Open pyVideoTrans
3. In the menu bar, select Speech Recognition (R) → Nvidia parakeet-tdt
4. In the configuration window, set the "HTTP address" to:

   http://127.0.0.1:5092/v1

5. Click "Save" to start using it

Parameter Reference

Parameter	Value	Notes
HTTP address	http://127.0.0.1:5092/v1	parakeet-api endpoint address
api_key	123456 (any value)	Fixed internally by the program — no need to change

Common Connection Errors

If pyVideoTrans reports a connection failure, check:

1. Is parakeet-api running? — Is the command line window still open?
2. Is the port correct? — Default port is 5092; ensure it's not occupied by another program
3. Address format — Must end with /v1, i.e., http://127.0.0.1:5092/v1
4. Firewall — Check if your local firewall is blocking port 5092

---

Source Code Deployment

If you prefer to deploy from source or need custom configuration, follow these steps.

Installation and Configuration Guide

This project supports Windows, macOS, and Linux. Follow the steps below to install and configure.

Step 0: Set Up Python 3.10 Environment

If you don't have Python 3 installed, follow this tutorial: https://pvt9.com/_posts/pythoninstall

Make sure to install Python 3.10 — other versions may have compatibility issues.

Step 1: Prepare FFmpeg

This project uses ffmpeg for audio/video format preprocessing.

• Windows (recommended):

  1. Download from the FFmpeg GitHub releases and extract ffmpeg.exe
  2. Place ffmpeg.exe directly in the project root directory (same level as app.py). The program will detect and use it automatically — no environment variable configuration needed.

• macOS (using Homebrew):

  brew install ffmpeg

• Linux (Debian/Ubuntu):

  sudo apt update && sudo apt install ffmpeg

Step 2: Create Python Virtual Environment and Install Dependencies

1. Download or clone the project code to your local machine (preferably in a non-system drive folder with an English or numeric name)

2. Open a terminal and navigate to the project root directory (on Windows, type cmd in the folder address bar and press Enter)

   

3. Create a virtual environment:

   python -m venv venv

4. Activate the virtual environment:

   • Windows (CMD/PowerShell): .\venv\Scripts\activate
   • macOS / Linux (Bash/Zsh): source venv/bin/activate

5. Install dependencies:

   • If you don't have an NVIDIA GPU (CPU only):

     pip install -r requirements.txt

   • If you have an NVIDIA GPU (GPU acceleration): a. Ensure the latest NVIDIA driver and matching CUDA Toolkit are installed b. Uninstall any existing old PyTorch: pip uninstall -y torch c. Install PyTorch matching your CUDA version (e.g., for CUDA 12.6):

     pip install torch --index-url https://download.pytorch.org/whl/cu126

Step 3: Start the Service

In a terminal with the virtual environment activated, run:

python app.py

First run will automatically download the model (~1.2 GB). Please wait patiently — the service will start automatically after the download.

If you see a stream of log messages, don't worry — just wait:

Startup success screen:

Seeing the interface above means the service started successfully.

Usage

Method 1: Web Interface

1. Open in browser: http://127.0.0.1:5092
2. Drag and drop or click to upload your audio/video file
3. Click "Start Transcription" and wait for processing to complete — the SRT subtitle will appear below and can be downloaded

Method 2: API Call (Python Example)

Using the openai library, you can easily call this service:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:5092/v1",
    api_key="any-key",
)

with open("your_audio.mp3", "rb") as audio_file:
    srt_result = client.audio.transcriptions.create(
        model="parakeet",
        file=audio_file,
        response_format="srt"
    )
print(srt_result)

Optimal Configuration

Setting	Recommended Value	Notes
Model	parakeet_srt_words	Default model, high accuracy
Compute device	GPU (CUDA)	Strongly recommended with NVIDIA GPU — tens of times faster
Audio format	WAV 16kHz mono	Best recognition results; other formats are auto-converted
Initial prompt	Domain-specific terms	Improves accuracy for specialized vocabulary

Common Issues

Q: "CUDA out of memory" error on startup?

A: Insufficient VRAM. Try:

• Close other programs using VRAM
• Use a smaller model
• Switch to CPU mode

Q: Timeline is misaligned in transcription results?

A: Check if the input audio has issues. Convert to standard WAV format (16kHz, mono) with FFmpeg before transcribing.

Q: Browser doesn't open automatically after starting the bundle?

A: Manually visit http://127.0.0.1:5092 in your browser. If it can't be accessed, check the command line window for error messages.

Q: Which languages are supported?

A: The Parakeet model primarily supports English. For Chinese or other languages, consider using other recognition channels (such as faster-whisper).