Audio Seek

A lightweight Python library for reading arbitrary time slices out of common audio files without loading them fully — with minimal dependencies (numpy + soundfile).

Key Features

• Zero Waste: Reads and decodes only the requested time slice, not the whole file.
• Usual Formats: Supports WAV, AIFF, FLAC, OGG/Vorbis, and MP3 via soundfile/libsndfile.
• O(1) Where Possible: Seeking into uncompressed PCM (WAV/AIFF) is a direct sample-offset seek, O(1). Compressed formats (MP3/FLAC/OGG) are frame/entropy coded, so seeking there is best-effort at frame granularity, not sample-exact.
• Complexity Introspection: AudioSeek.seek_complexity() tells you which regime a given file falls into.
• Minimal Dependencies: Just numpy and soundfile — no bundled codecs, no JIT toolchains.

Installation

pip install audio-seek

Quick Start

Read a Specific Audio Segment

from audio_seek import read_audio_segment

# Read 5 seconds starting at 2 minutes
segment = read_audio_segment(
    file_path="long_audio.wav",
    start_sec=120.0,
    duration_sec=5.0,
)
# Returns a numpy float32 array of shape (sample_rate * duration_sec,)

Get Duration Without Loading Audio

from audio_seek import AudioSeek

# Only reads the header, doesn't decode audio data
duration = AudioSeek.get_duration("audio.mp3")
print(f"Duration: {duration:.2f}s")

Check Seek Complexity

from audio_seek import AudioSeek

AudioSeek.seek_complexity("audio.wav")  # "O(1)"  — uncompressed PCM
AudioSeek.seek_complexity("audio.mp3")  # "O(n)"  — frame/entropy coded

Extract a Segment to a New WAV File

from audio_seek import AudioSeek

AudioSeek.read_segment_to_file(
    file_path="input.mp3",
    start_sec=10.0,
    duration_sec=15.0,
    output_path="clip.wav",  # written as PCM_16 WAV
)

Mix Multi-Channel Audio Down to Mono

from audio_seek import ensure_mono

mono = ensure_mono(stereo_data, style="soundfile")  # (samples, channels) -> (samples,)

How It Works

audio-seek wraps soundfile/libsndfile and seeks directly via SoundFile.seek():

• For uncompressed PCM (WAV, AIFF), each sample occupies a fixed number of bytes, so seeking is a direct file-offset computation — true O(1), sample-accurate.
• For compressed formats (MP3, FLAC, OGG/Vorbis), audio is encoded in frames/blocks of varying or fixed sample counts, sometimes with decoder warm-up requirements. Seeking there lands on the nearest frame boundary and is not guaranteed sample-exact — closer to O(n) in the worst case.

Use Cases

Ideal for applications requiring low-latency access to specific segments of long audio recordings:

• Machine Learning: Dataset slicing without loading entire files
• Web Services: Real-time audio segment delivery
• Audio Processing: Random access pipelines

Requirements

• Python >= 3.11
• numpy
• soundfile

Testing

pytest tests/ -v

License

Apache License 2.0 - see LICENSE file for details.