AvioFlow

AvioFlow is a high-performance and easy-to-use streaming audio decoding library.

The avioflow project is build on top of the FFMPEG library.

Features

• Audio format: mp3, opus, flac, ogg, wav, m4a, aac. Anything FFmpeg supports — see the Supported Formats Reference for the full decoder/encoder list.
• Flexible Input: Files, URLs, memory buffers, and real-time streams
• Hardware Capture: WASAPI loopback (system audio) and DirectShow (microphones)
• Resampling: Built-in sample rate conversion
• Zero-copy API: Direct buffer access via FrameData for maximum performance
• Cross-platform: Windows, Linux, macOS

Speed

Decode time (best of 10 runs) for public/wavs/TownTheme.mp3 (MP3, 44.1kHz stereo, ~97.5s) on a single machine, decoding the full file into memory both without resampling and with resampling to 16kHz. Numbers are for illustration on this environment, not a formal cross-platform benchmark.

Supported language

AvioFlow is packaged for several runtime and application environments. The native core is shared across bindings, so behavior stays consistent whether you embed it in a C++ service, call it from Python or JavaScript, ship it in a JVM application, or run it in WebAssembly.

Decoder API Flow

AvioFlow uses the same pull-style output functions for offline and streaming decoding. The difference is only how input bytes enter the decoder.

Offline Input

+-----------------------+
| AudioDecoder(options) |
+-----------+-----------+
            |
            v
+-----------------------------+
| load_file(path)             |
| load_buffer(bytes, size)    |
+-----------+-----------------+
            |
            v
+-----------------------------+
| get_frame()                          |  one decoded frame, zero-copy
| get_samples(start, stop)             |  samples in [start, stop), defaults to all
+-----------+-----------------+
            |
            v
+-----------------------------+
| is_finished()               |
+-----------------------------+

get_samples(start_seconds, stop_seconds) supports offline seek/time-range decoding: pass a half-open range in seconds to decode only that window, or call it with no arguments to decode the whole file. It may be called multiple times on the same decoder to fetch different ranges; each call seeks independently.

Streaming Input

+--------------------------------------+
| AudioDecoder(input_format, rate, ch) |
+-----------+--------------------------+
            |
            v
+-----------------------------+
| feed(chunk)                 |  first feed starts stream mode
+-----------+-----------------+
            |
            v
+-----------------------------+
| get_frame()                 |  returns empty if data is incomplete
| get_samples()               |  drains currently available samples (start/stop range not supported here)
+-----------+-----------------+
            | repeat feed/get_* while streaming
            v
+-----------------------------+
| flush()                     |  no more input; drain decoder delay
+-----------+-----------------+
            |
            v
+-----------------------------+
| get_samples() / get_frame() |  drain until is_finished()
+-----------------------------+

flush() does not discard data. It marks stream input complete so remaining buffered bytes and codec-delayed frames can be drained.

Build from Source

Prerequisites

• CMake 3.20+
• Visual Studio 2022+ (Windows) or GCC 11+ (Linux)

C++ Build

cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON
cmake --build build --config Release

FFmpeg is fetched and configured automatically during the CMake configure step.

Run the tests:

ctest --test-dir build --output-on-failure

Installation

Python

pip install avioflow

Java

Gradle users need the main Java API jar plus one native classifier for the target platform:

dependencies {
    implementation("io.github.lxp3:avioflow:0.7.3")
    runtimeOnly("io.github.lxp3:avioflow:0.7.3:linux-x86_64")
}

Maven:

<dependency>
  <groupId>io.github.lxp3</groupId>
  <artifactId>avioflow</artifactId>
  <version>0.3.2</version>
</dependency>
<dependency>
  <groupId>io.github.lxp3</groupId>
  <artifactId>avioflow</artifactId>
  <version>0.3.2</version>
  <classifier>linux-x86_64</classifier>
  <scope>runtime</scope>
</dependency>

Native classifiers: linux-x86_64, linux-aarch64, macos-x86_64, macos-aarch64, windows-x86_64, windows-aarch64.

C++ (CMake)

Download the C++ package for your platform and linkage, then point CMake at the extracted package root with CMAKE_PREFIX_PATH.

find_package(avioflow CONFIG REQUIRED)
target_link_libraries(your_target PRIVATE avioflow::avioflow)

Release packages are split by linkage and platform:

Linux binaries target glibc 2.28 or newer. Select ABI 0 for the legacy libstdc++ string ABI or ABI 1 for the C++11 string ABI; the eight Linux C++ archives use the filenames avioflow-{shared|static}-linux-{x64|arm64}-abi{0|1}.tar.gz.

• avioflow-shared-linux-x64-abi1, avioflow-shared-linux-x64-abi0
• avioflow-static-linux-x64-abi1, avioflow-static-linux-x64-abi0
• avioflow-shared-linux-arm64-abi1, avioflow-shared-linux-arm64-abi0
• avioflow-static-linux-arm64-abi1, avioflow-static-linux-arm64-abi0
• avioflow-shared-macos-x64, avioflow-static-macos-x64
• avioflow-shared-macos-arm64, avioflow-static-macos-arm64
• avioflow-shared-win-x64, avioflow-static-win-x64
• avioflow-shared-win-arm64, avioflow-static-win-arm64

Shared packages include the FFmpeg dynamic libraries needed at runtime. Static packages include FFmpeg static libraries, transitive static dependency metadata, and the bundled FFmpeg CMake package, so consumers do not need to configure FFmpeg separately.

C++ API

Core Classes

AudioDecoder

Main class for audio decoding.

#include "avioflow-cxx-api.h"
using namespace avioflow;

// Constructor options
AudioStreamOptions options;
options.output_sample_rate = 16000;    // Target sample rate
options.input_format = "s16le";        // For streaming: source format
options.input_sample_rate = 48000;     // For streaming: source rate
options.input_channels = 2;            // For streaming: source channels

AudioDecoder decoder(options);

Methods

FrameData

Zero-copy frame data structure returned by get_frame().

struct FrameData {
    float** data;        // Planar channel pointers: data[channel][sample]
    int num_channels;    // Number of channels
    int num_samples;     // Samples per channel

    operator bool();     // True if valid data
};

⚠️ Warning: FrameData.data points to internal buffer, valid only until next get_frame() or get_samples() call.

Examples

File Decoding (Offline)

AudioDecoder decoder({.output_sample_rate = 16000});
decoder.load_file("audio.mp3");

auto samples = decoder.get_samples();  // vector<vector<float>>
std::cout << "Channels: " << samples.size() << std::endl;
std::cout << "Samples: " << samples[0].size() << std::endl;

Frame-by-Frame Decoding

AudioDecoder decoder;
decoder.load_file("audio.mp3");

while (auto frame = decoder.get_frame()) {
    // frame.data[channel][sample]
    for (int c = 0; c < frame.num_channels; c++) {
        process(frame.data[c], frame.num_samples);
    }
}

Raw PCM Memory Decode

// Raw PCM bytes have no container/header, so provide the input format details.
// Use FFmpeg demuxer format names such as "s16le", not codec names like
// "pcm_s16le".
AudioStreamOptions opts;
opts.input_format = "s16le";       // Signed 16-bit little-endian PCM
opts.input_sample_rate = 8000;     // 8 kHz
opts.input_channels = 1;           // Mono

AudioDecoder decoder(opts);
decoder.load_buffer(pcm_bytes, pcm_size); // Full PCM buffer in memory

while (auto frame = decoder.get_frame()) {
    // Output samples are float planar: frame.data[channel][sample]
    process(frame.data[0], frame.num_samples);
}

Time-Range Decoding (Offline Seek)

AudioDecoder decoder;
decoder.load_file("audio.mp3");

// Decode only seconds 10.3 to 20.3
auto samples = decoder.get_samples(10.3, 20.3);

// Can be called again with a different range on the same decoder
auto next_range = decoder.get_samples(30.0, 40.0);

Streaming Decode (Push-based)

AudioStreamOptions opts;
opts.input_format = "s16le";
opts.input_sample_rate = 48000;
opts.input_channels = 2;

AudioDecoder decoder(opts);
decoder.feed(raw_bytes, size);  // First feed starts stream mode

auto samples = decoder.get_samples(); // Decode all buffered data
// Or frame-by-frame:
while (auto frame = decoder.get_frame()) {
    // Process decoded audio...
}
decoder.flush();

Python API

AudioDecoder

import avioflow

# Constructor with keyword arguments
decoder = avioflow.AudioDecoder(
    output_sample_rate=16000,    # Optional: target sample rate
    input_format="s16le",        # For streaming: source format
    input_sample_rate=48000,     # For streaming: source rate
    input_channels=2             # For streaming: source channels
)

Methods

Metadata

# Quick metadata inspection without full decoding
meta = avioflow.info("audio.mp3")
print(f"Duration: {meta.duration}s")
print(f"Sample Rate: {meta.sample_rate}Hz")
print(f"Codec: {meta.codec}")

# Encoded audio bytes also work
with open("audio.mp3", "rb") as f:
    meta = avioflow.info(f.read())

Examples

File Decoding

decoder = avioflow.AudioDecoder(output_sample_rate=16000)
meta = decoder.load_file("speech.wav")
samples = decoder.get_samples()      # numpy array (channels, samples)
print(f"Shape: {samples.shape}")     # e.g., (1, 160000)

Time-Range Decoding (Offline Seek)

decoder = avioflow.AudioDecoder()
decoder.load_file("audio.mp3")

# Decode only seconds 10.3 to 20.3
samples = decoder.get_samples(10.3, 20.3)

# Can be called again with a different range on the same decoder
next_range = decoder.get_samples(30.0, 40.0)

Streaming Decode

decoder = avioflow.AudioDecoder(
    input_format="s16le",
    input_sample_rate=48000,
    input_channels=2
)

while True:
    data = socket.recv(4096)
    if not data:
        decoder.flush()
        break
    decoder.feed(data)
    samples = decoder.get_samples()
    if samples.size > 0:
        process_audio(samples)

Device Discovery

devices = avioflow.DeviceManager.list_audio_devices()
for dev in devices:
    print(f"{dev.name}: {dev.description}")

Logging

avioflow.set_log_level("debug")  # quiet, error, warning, info, debug, trace

Node.js API

Compatibility

Installation

npm install avioflow

ESM Import

import avioflow from 'avioflow';

Module-level Functions

AudioDecoder

// Constructor with options object
const decoder = new avioflow.AudioDecoder({
    outputSampleRate: 16000,    // Optional: target sample rate
    outputNumChannels: 1,       // Optional: target channels
    inputFormat: 's16le',       // For streaming: source format
    inputSampleRate: 48000,     // For streaming: source rate
    inputChannels: 2            // For streaming: source channels
});

Methods

Examples

Quick File Loading (Recommended)

// Opens file, resamples to 16kHz mono, and decodes everything
const { metadata, samples } = avioflow.load("audio.mp3", {
    outputSampleRate: 16000,
    outputNumChannels: 1
});

console.log(`Duration: ${metadata.duration}s`);
console.log(`Channels: ${samples.length}, Samples: ${samples[0].length}`);

Batch Decoding with Decoder Instance

const decoder = new avioflow.AudioDecoder({ outputSampleRate: 44100 });
const meta = decoder.loadFile("audio.wav");

// Decodes the entire file into memory
const allSamples = decoder.getSamples();
process(allSamples);

Time-Range Decoding (Offline Seek)

const decoder = new avioflow.AudioDecoder();
decoder.loadFile("audio.mp3");

// Decode only seconds 10.3 to 20.3
const samples = decoder.getSamples(10.3, 20.3);

// Can be called again with a different range on the same decoder
const nextRange = decoder.getSamples(30.0, 40.0);

Streaming Decode (Real-time)

const decoder = new avioflow.AudioDecoder({
    inputFormat: 's16le',
    inputSampleRate: 48000,
    inputChannels: 2
});

socket.on('data', (chunk) => {
    decoder.feed(chunk);

    // Get all samples decoded from this chunk
    const samples = decoder.getSamples();
    if (samples.length > 0) {
        processAudio(samples);
    }
});

socket.on('end', () => {
    decoder.flush();
    const remaining = decoder.getSamples();
    if (remaining.length > 0) {
        processAudio(remaining);
    }
});

Device Discovery

const devices = avioflow.listAudioDevices();
devices.forEach(dev => {
    console.log(`${dev.isOutput ? 'Output' : 'Input'}: ${dev.name} (${dev.description})`);
});

Java API

File Decoding

import io.github.lxp3.avioflow.AudioDecoder;
import io.github.lxp3.avioflow.AudioStreamOptions;

try (AudioDecoder decoder = new AudioDecoder(
        new AudioStreamOptions().outputSampleRate(16000))) {
    decoder.loadFile("audio.mp3");
    float[][] samples = decoder.getSamples();
    System.out.println(samples.length + " channels");

    // Decode only seconds 10.3 to 20.3 (offline mode)
    float[][] range = decoder.getSamples(10.3, 20.3);
}

Encoding

import io.github.lxp3.avioflow.AudioEncoder;
import io.github.lxp3.avioflow.AudioWriteOptions;

AudioEncoder.saveAudio(
    "out.wav",
    samples,
    new AudioWriteOptions()
        .containerFormat("wav")
        .codecName("pcm_s16le")
        .sampleRate(16000)
);

Rust API

Full reference: rust/README.md.

File Decoding

use avioflow::{AudioDecoder, StreamOptions};

let mut decoder = AudioDecoder::new(&StreamOptions::new().output_sample_rate(16000))?;
let metadata = decoder.load_file("audio.mp3")?;
let samples = decoder.get_samples()?;
println!("{} channels at {} Hz", samples.len(), metadata.sample_rate);

// Decode only seconds 10.3 to 20.3
let range = decoder.get_samples_range(10.3, Some(20.3))?;

Resampling and Encoding

use avioflow::{resample, save_audio, WriteOptions};

let mono_16k = resample(&samples, 44100, 16000, Some(1))?;

save_audio("out.wav", &mono_16k, &WriteOptions::new()
    .container_format("wav")
    .codec_name("pcm_s16le")
    .sample_rate(16000))?;

License

MIT License