qvac-lib-infer-whispercpp

This library simplifies running inference with the Whisper transcription model within QVAC runtime applications. It provides an easy interface to load, execute, and manage Whisper inference instances, supporting multiple data sources (data loaders).

Note: This library now uses whisper.cpp for improved performance and compatibility. The previous MLC-based implementation has been replaced.

Table of Contents

• Supported Platforms
• Installation
• Development
• Usage
  • 1. Choose a Data Loader
  • 2. Configure Transcription Parameters
  • 3. Define Model Files Configuration
  • 4. Create Model Instance
  • 5. Load Model
  • 6. Run Transcription
  • 7. Release Resources
• Quickstart example
• Model registry
• Other examples
• Resources
• License

Supported Platforms

Platform	Architecture	Min Version	Status	GPU Support
macOS	arm64, x64	14.0+	✅ Tier 1	Metal
iOS	arm64	17.0+	✅ Tier 1	Metal
Linux	arm64, x64	Ubuntu-22+	✅ Tier 1	Vulkan
Android	arm64	12+	✅ Tier 1	Vulkan
Windows	x64	10+	✅ Tier 1	Vulkan

Dependencies:

• qvac-lib-inference-addon-cpp (=0.12.2): C++ addon framework
• qvac-fabric-whisper.cpp (latest): Inference engine
• Bare Runtime (≥1.24.2): JavaScript runtime
• Linux requires Clang/LLVM 19 with libc++

Installation

Prerequisites

Make sure Bare Runtime is installed:

npm install -g bare bare-make

Note : Make sure the Bare version is >= 1.24.2. Check this using :

bare -v

Installing the Package

Install the latest development version (adjust package name based on desired model/quantization):

npm install @qvac/transcription-whispercpp@latest

Development

Building the AddOn Locally

For local development, you'll need to build the native addon that interfaces with the Whisper model. Follow these steps:

Prerequisites

First, make sure you have the prerequisites installed as described in the Installation section.

System Requirements

Supported Platforms:

• Linux (x64, ARM64)
• macOS (x64, ARM64)
• Windows (x64)

Required Tools

All Platforms:

• CMake (>= 3.25)
• Git with submodule support
• C++ Compiler with C++20 support
  • Linux: GCC 9+ or Clang 10+
  • macOS: Xcode 12+ (provides Clang 12+)
  • Windows: Visual Studio 2019+ or MinGW-w64

vcpkg Setup

This project uses vcpkg for C++ dependency management. You need to install and configure vcpkg before building:

1. Install vcpkg:

# Clone vcpkg
git clone https://github.com/Microsoft/vcpkg.git
cd vcpkg

# Bootstrap vcpkg
# On Linux/macOS:
./bootstrap-vcpkg.sh
# On Windows:
.\bootstrap-vcpkg.bat

2. Set Environment Variable:

# Linux/macOS (add to your shell profile):
export VCPKG_ROOT=/path/to/vcpkg

# Windows (PowerShell):
$env:VCPKG_ROOT = "C:\path\to\vcpkg"
# Or set permanently via System Properties > Environment Variables

3. Integrate vcpkg (optional but recommended):

# This makes vcpkg available to all CMake projects
./vcpkg integrate install

Platform-Specific Setup

Linux:

# Ubuntu/Debian
sudo apt update
sudo apt install build-essential cmake git pkg-config

# CentOS/RHEL/Fedora
sudo yum groupinstall "Development Tools"
sudo yum install cmake git pkgconfig
# or for newer versions:
sudo dnf groupinstall "Development Tools"
sudo dnf install cmake git pkgconfig

macOS:

# Install Xcode Command Line Tools
xcode-select --install

# Using Homebrew (recommended)
brew install cmake git

# Using MacPorts
sudo port install cmake git

Windows:

• Install Visual Studio 2019+ with C++ development tools
• Or install Build Tools for Visual Studio
• Install CMake (3.25+)
• Install Git for Windows

GPU Acceleration (Optional)

Metal (macOS):

• Automatically available on macOS 10.13+ with Metal-capable hardware
• No additional setup required

Vulkan (Linux/Windows):

• Install Vulkan SDK from LunarG
• Ensure your GPU drivers support Vulkan 1.1+

Linux Vulkan Setup:

# Ubuntu/Debian
sudo apt install vulkan-tools libvulkan-dev vulkan-utility-libraries-dev spirv-tools

# CentOS/RHEL/Fedora
sudo yum install vulkan-tools vulkan-devel vulkan-validation-layers-devel spirv-tools

Windows Vulkan Setup:

• Download and install the Vulkan SDK
• Ensure your graphics drivers are up to date

Clone and Setup

git clone https://github.com/tetherto/qvac.git
cd qvac/packages/qvac-lib-infer-whispercpp

# Initialize submodules (required for native dependencies)
git submodule update --init --recursive

# Install dependencies
npm install

Build the Native AddOn

# Build the native addon
npm run build

This command runs the complete build sequence:

1. bare-make generate - Generates build configuration
2. bare-make build - Compiles the native C++ addon
3. bare-make install - Installs the built addon

Running Tests

After building, you can run the test suite:

# Run all tests (lint + unit + integration)
npm test

# Or run tests individually
npm run test:unit
npm run test:integration

Integration tests cover both the chunked reload flow (test/integration/audio-ctx-chunking.test.js) and the live streaming flow (test/integration/live-stream-simulation.test.js), so running them is the quickest way to verify those end-to-end scenarios after changes.

Development Workflow

For ongoing development, the typical workflow is:

npm install && npm run build && npm run test:integration

Usage

The library provides a straightforward workflow for audio transcription:

Heads up: the package is intended to be used through index.js’s TranscriptionWhispercpp class. Advanced sections below document the native addon for completeness, but you rarely need them when integrating the published npm package.

1. Choose a Data Loader

Data loaders abstract the way model files are accessed, whether from the filesystem, a network drive, or any other storage mechanism. More info about model registry and model builds in resources.

• Filesystem Data Loader

First, select and instantiate a data loader that provides access to model files:

const FilesystemDL = require('@qvac/dl-filesystem')
const fsDL = new FilesystemDL({
  dirPath: './path/to/model/files' // Directory containing model weights and settings
})

2. Configure Transcription Parameters

Most users interact with the addon exclusively through index.js. From that entrypoint we surface a small, safe subset of options; everything else keeps whisper.cpp defaults.

What index.js accepts

Section	Key	Description
contextParams	model	Absolute or relative path to the .bin whisper model
		(all other context keys keep their defaults because changing them forces a full reload, see below)
whisperConfig	(any whisper_full_params key)	Forwarded untouched. We surface convenience defaults in index.js, but every whisper.cpp flag is accepted—see Advanced configuration.
miscConfig	caption_enabled	Formats segments with <|start|>..<|end|> markers

Context keys that force a full reload

Internally WhisperModel::configContextIsChanged() watches model, use_gpu, flash_attn and gpu_device. If any of these change we must:

1. Call unload() (destroys the current whisper_context and whisper_state).
2. Recreate the context via whisper_init_from_file_with_params.
3. Warm up the model again before the next job.

Depending on model size this can take several seconds. Everything else in whisperConfig—language, temperatures, VAD settings, etc.—is applied in place and does not trigger a reload. If you are seeing unexpected pauses, double-check that you are not mutating these four context keys between jobs.

Advanced configuration

Need more than the handful of options exposed in index.js? The upstream whisper.cpp documentation lists every flag available through whisper_full_params. Rather than duplicating that matrix here, refer to:

• The official parameter reference: whisper_full_params
• Our longer examples for concrete shapes:
  • examples/example.audio-ctx-chunking.js (shows offset_ms, duration_ms, audio_ctx, and reload loops)
  • examples/example.live-transcription.js (shows streaming chunks into a single job)

Those scripts stay in sync with the codebase and are the best place to copy from when you need the raw addon surface.

3. Configuration Example

Quick JS-level configuration (what you typically pass to new TranscriptionWhispercpp(...)):

const config = {
  contextParams: {
    model: './models/ggml-tiny.bin'
  },
  whisperConfig: {
    language: 'en',
    duration_ms: 0,
    temperature: 0.0,
    suppress_nst: true,
    n_threads: 0,
    vad_model_path: './models/ggml-silero-v5.1.2.bin',
    vadParams: {
      threshold: 0.6,
      min_speech_duration_ms: 250,
      min_silence_duration_ms: 200
    }
  },
  miscConfig: {
    caption_enabled: false
  }
}

Between this minimal configuration and the example scripts you should have everything needed, whether you are wiring the addon by hand or just instantiating TranscriptionWhispercpp.

Available Whisper Models (from HuggingFace):

Model	Size	Description
ggml-tiny.bin	78 MB	Smallest, fastest
ggml-base.bin	148 MB	Balanced size/accuracy
ggml-small.bin	488 MB	Better accuracy
ggml-medium.bin	1.5 GB	High accuracy
ggml-large-v3.bin	3.1 GB	Best accuracy
ggml-large-v3-turbo.bin	1.6 GB	Best accuracy, faster

Quantized variants (q8_0, q5_1, q5_0) are also available for all sizes.

VAD Model (from HuggingFace):

Model	Size	Description
ggml-silero-v5.1.2.bin	885 KB	Silero VAD for voice activity detection

Downloading Models

Use the provided script to download models from HuggingFace:

npm run download-models

Or download manually with curl:

# Whisper model
curl -L -o models/ggml-tiny.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin

# VAD model
curl -L -o models/ggml-silero-v5.1.2.bin https://huggingface.co/ggml-org/whisper-vad/resolve/main/ggml-silero-v5.1.2.bin

4. Create Model Instance

Import the specific Whisper model class based on the installed package and instantiate it:

const TranscriptionWhispercpp = require('@qvac/transcription-whispercpp')

const model = new TranscriptionWhispercpp(args, config)

Note : This import changes depending on the package installed.

5. Load Model

Load the model weights and initialize the inference engine. Optionally provide a callback for progress updates:

try {
  // Basic usage
  await model.load()

  // Advanced usage with progress tracking
  await model.load(
          false,  // Don't close loader after loading
          (progress) => console.log(`Loading: ${progress.overallProgress}% complete`)
  )
} catch (error) {
  console.error('Failed to load model:', error)
}

Progress Callback Data

The progress callback receives an object with the following properties:

Property	Type	Description
action	string	Current operation being performed
totalSize	number	Total bytes to be loaded
totalFiles	number	Total number of files to process
filesProcessed	number	Number of files completed so far
currentFile	string	Name of file currently being processed
currentFileProgress	string	Percentage progress on current file
overallProgress	string	Overall loading progress percentage

6. Run Transcription

Pass an audio stream (e.g., from bare-fs.createReadStream) to the run method. Process the transcription results asynchronously.

There are two ways to receive transcription results:

Option 1: Real-time Streaming with onUpdate()

The onUpdate() callback receives each transcription segment in real-time as whisper.cpp generates them during processing. This is ideal for live transcription display or progressive updates.

try {
  const audioStream = fs.createReadStream('path/to/your/audio.ogg', {
    highWaterMark: 16000 // Adjust based on bitrate (e.g., 128000 / 8)
  })

  const response = await model.run(audioStream)

  // Receive segments as they are transcribed (real-time streaming)
  await response
          .onUpdate(segment => {
            console.log('New segment transcribed:', segment)
            // Each segment arrives immediately after whisper.cpp processes it
          })
          .await() // Wait for transcription to complete

  console.log('Transcription finished!')

} catch (error) {
  console.error('Transcription failed:', error)
}

Option 2: Complete Result with iterate()

The iterate() method returns all transcription segments after the entire transcription completes. This is useful when you need the full result before processing.

try {
  const audioStream = fs.createReadStream('path/to/your/audio.ogg', {
    highWaterMark: 16000
  })

  const response = await model.run(audioStream)

  // Wait for complete transcription, then iterate over all segments
  for await (const transcriptionChunk of response.iterate()) {
    console.log('Transcription chunk:', transcriptionChunk)
  }

  console.log('Transcription finished!')

} catch (error) {
  console.error('Transcription failed:', error)
}

Key Differences:

• onUpdate(): Real-time streaming - segments arrive as they are generated by whisper.cpp's new_segment_callback
• iterate(): Batch processing - all segments available after transcription completes

Chunking long recordings with reload()

examples/example.audio-ctx-chunking.js shows the production pattern: reuse a model instance, call reload() with { offset_ms, duration_ms, audio_ctx } per chunk (first chunk uses audio_ctx = 0, subsequent ones clamp to ~1500), then run the full audio stream. The matching integration test (test/integration/audio-ctx-chunking.test.js) exercises exactly the same flow.

Live streaming a single job

examples/example.live-transcription.js feeds tiny PCM buffers into a pushable Readable, keeps a single model.run(...) open, and relies on onUpdate() for incremental text. test/integration/live-stream-simulation.test.js covers both the streaming case and a segmented loop without any reload() calls.

7. Release Resources

Always unload the model when finished to free up memory and resources:

try {
  await model.unload()
} catch (error) {
  console.error('Failed to unload model:', error)
}

Quickstart example

1. Clone the repo & Install the dependencies

git clone https://github.com/tetherto/qvac-lib-infer-whispercpp.git
cd qvac-lib-infer-whispercpp
npm install

2. Run the quickstart example inside examples folder

bare examples/quickstart.js

3. Code Walkthrough

See examples/quickstart.js for the full workflow (TranscriptionWhispercpp + filesystem loader), including streaming audio and cleanup.

Benchmarking

We conduct comprehensive benchmarking of our Whisper transcription models to evaluate their performance across different audio conditions and metrics. The evaluations are performed using the LibriSpeech dataset, a standard benchmark for speech recognition tasks.

Our benchmarking suite measures transcription accuracy using Word Error Rate (WER) and Character Error Rate (CER), along with performance metrics such as model load times and inference speeds.

Benchmark Results

For detailed benchmark results across all supported audio conditions and model configurations, see our Benchmark Results Summary.

The benchmarking covers:

• Transcription Accuracy: WER and CER scores for accuracy assessment
• Performance Metrics: Model loading times and inference speeds
• Audio Conditions: Clean speech, noisy environments and varied accents
• Model Variants: Different quantization levels and model sizes

Results are updated regularly as new model versions are released.

Other examples

• Quickstart – Basic transcription example.
• Standalone Decoder – Demonstrates the FFmpeg decoder independently for audio format conversion.
• Model Reload – Shows how to reload models with different configurations (language, temperature).
• Audio ctx chunking – Processes long recordings by reloading with offset_ms, duration_ms, and audio_ctx per chunk (mirrors the audio-ctx-chunking integration test).
• Live transcription – Streams small chunks into a single job while maintaining Whisper state between updates (mirrors the live-stream-simulation integration test).

Glossary

• Bare – Small and modular JavaScript runtime for desktop and mobile. Learn more.
• QVAC – QVAC is our open-source AI-SDK for building decentralized AI applications.
• Corestore – Corestore is a Hypercore factory that makes it easier to manage large collections of named Hypercores. Learn more.

Error Range

All the errors from this library are in the range of 6001-7000

Resources

• PoC Repo: tetherto/qvac-transcription-poc
• Pear app (Desktop): TBD

License

This project is licensed under the Apache-2.0 License – see the LICENSE file for details.

For questions or issues, please open an issue on the GitHub repository.