audiopy_ai 🎧

Open-Source Framework for Intelligent Audio Editing (DSP + AI) — inspired by OpenCV, but designed for audio.

audiopy_ai is a unified Python library that seamlessly bridges the gap between traditional Digital Signal Processing (DSP) and modern deep-learning-based AI analysis. It enables developers to load, manipulate, filter, stretch, and analyze audio files via a clean, high-performance, and consistent API.

---

✨ Features

• 📂 High-Performance I/O: Fast loading and saving of audio files (WAV, MP3, FLAC, etc.) into unified numpy arrays.
• 🎚️ DSP Audio Editing: Trimming, fading, peak-normalization, gain adjustment, reversing, and sample concatenation.
• 🔊 Advanced DSP Effects:
  • Butterworth low-pass and high-pass filters (via Scipy SOS filters).
  • Convolution-based reverb with impulse response normalization.
  • Resampling-based time-stretching.
• 🤖 AI-Powered Analysis:
  • Speech-to-Text transcription (using Whisper).
  • Audio classification and sound event tagging (using HuBERT / AudioSet models).
• 🔌 Zero-Config FFmpeg: Automated local path auto-discovery for FFmpeg (supporting Scoop, Winget, and custom Program Files locations) to ensure seamless audio decoding out-of-the-box on Windows and Unix systems.

---

📦 Installation

1. Core Install (DSP & I/O)

pip install audiopy_ai

2. Advanced Install (With Librosa for Feature Extraction)

pip install audiopy_ai[advanced]

⚡ System Requirements

• Python: >= 3.9
• FFmpeg: Required for decoding compressed formats (MP3, M4A, etc.) and video analysis. audiopy_ai automatically searches standard environment directories, Winget packages, and Scoop shims to configure FFmpeg on import.

---

🚀 Quick Start

Here is a simple example showing how to load a file, normalize it, apply a fade, and run AI-based classification on it:

import audiopy_ai as ap

# 1. Load audio file (WAV, MP3, etc.)
samples, sr = ap.load("input.wav")
print(f"Loaded {len(samples)} samples at {sr}Hz")

# 2. DSP Processing: Normalize amplitude & apply fade
normalized = ap.normalize(samples, target_db=-15.0)
processed = ap.fade(normalized, sr, fade_in_s=0.5, fade_out_s=0.5)

# 3. Save processed file
ap.save("output_processed.wav", processed, sr)
print("Saved processed audio!")

# 4. Run AI Event Analysis
extractor = ap.SoundComponentExtractor(analysis_type="audio_tagging")
analysis = extractor.extract_components("output_processed.wav")

print("Detected Components:")
for item in analysis.get("results", []):
    print(f" • {item['label']}: {item['score']:.1%}")

---

📖 API Reference

🔊 I/O Module

ap.load(path: str, mono: bool = True) -> (np.ndarray, int)

Loads an audio file into a 1D float32 numpy array. If mono=True and the file is stereo, channels are automatically averaged.

samples, sr = ap.load("sound.mp3")

ap.save(path: str, samples: np.ndarray, sr: int) -> None

Saves a numpy array back into a sound file.

ap.save("output.wav", samples, sr)

---

⚙️ DSP Editing Module

ap.trim(samples: np.ndarray, sr: int, start_s: float, end_s: float) -> np.ndarray

Trims audio to a specified time window (in seconds).

trimmed = ap.trim(samples, sr, start_s=1.0, end_s=5.0)

ap.concat(list_of_samples: list[np.ndarray]) -> np.ndarray

Concatenates multiple audio sample arrays sequentially.

full_audio = ap.concat([clip1, clip2])

ap.apply_gain(samples: np.ndarray, db: float) -> np.ndarray

Applies a gain adjustment (in decibels) to the audio samples.

louder = ap.apply_gain(samples, db=6.0)

ap.normalize(samples: np.ndarray, target_db: float = -20.0) -> np.ndarray

Normalizes the audio samples to match a target RMS decibel level.

normalized = ap.normalize(samples, target_db=-18.0)

ap.fade(samples: np.ndarray, sr: int, fade_in_s: float = 0.5, fade_out_s: float = 0.5) -> np.ndarray

Applies linear fade-in and fade-out envelopes to the audio.

faded = ap.fade(samples, sr, fade_in_s=1.0, fade_out_s=1.0)

ap.reverse(samples: np.ndarray) -> np.ndarray

Reverses the audio array (plays backwards).

reversed_audio = ap.reverse(samples)

ap.rms_db(samples: np.ndarray, eps: float = 1e-9) -> float

Calculates the Root-Mean-Square (RMS) volume of the audio in decibels.

volume_db = ap.rms_db(samples)

---

🎚️ Effects Submodule

Filters (audiopy_ai.effects.filters)

• butter_lowpass(sr, cutoff_hz, order=4): Generates low-pass filter coefficients.
• butter_highpass(sr, cutoff_hz, order=4): Generates high-pass filter coefficients.
• apply_sos_filter(samples, sos): Applies the generated filter coefficients (SOS format) to the audio.

from audiopy_ai.effects.filters import butter_lowpass, apply_sos_filter

# Filter high frequencies out above 1000Hz
sos = butter_lowpass(sr, cutoff_hz=1000)
filtered = apply_sos_filter(samples, sos)

Reverb (audiopy_ai.effects.reverb)

• reverb(samples, ir, wet=0.3): Applies a convolution-based reverb effect using an Impulse Response (IR) audio array.

from audiopy_ai.effects.reverb import reverb

# ir_samples is another audio array loaded from a reverb room impulse response file
wet_audio = reverb(samples, ir_samples, wet=0.4)

Time stretching (audiopy_ai.effects.time)

• time_stretch(samples, rate): Changes the playback speed of the audio without pitch shift (using resampling). rate > 1.0 speeds it up, while rate < 1.0 slows it down.

from audiopy_ai.effects.time import time_stretch

fast_audio = time_stretch(samples, rate=1.5)

---

🤖 AI-Powered Analysis Module

ap.AudioAnalyzer(model_name: str = "facebook/wav2vec2-base", api_key: str = None)

Initializes the HuggingFace-backed pipeline analyzer.

• Speech-to-Text models: openai/whisper-base, facebook/wav2vec2-base
• Audio Tagging models: superb/hubert-base-superb-ks (keyword spotting)

from audiopy_ai.ai_analysis import AudioAnalyzer

analyzer = AudioAnalyzer(model_name="openai/whisper-base")
results = analyzer.analyze_audio("speech.wav")
print("Transcribed Text:", results["components"]["metadata"]["transcription"])

ap.SoundComponentExtractor(analysis_type: str = "audio_tagging", api_key: str = None)

Simplified helper wrapper for executing AI analyses.

from audiopy_ai.ai_analysis import SoundComponentExtractor

# Instantiate for Sound Event detection
extractor = SoundComponentExtractor(analysis_type="audio_tagging")
info = extractor.extract_components("recording.wav")

---

🤝 Contributing

Contributions are always welcome!

1. Fork the repository on GitHub.
2. Install the library locally in editable mode:

   pip install -e .
   

3. Run the unit tests to ensure everything is working:

   pytest tests/
   

4. Open a Pull Request with your feature branch.

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.