AutoWave 2.0.0

The simplest way to classify audio in Python — pretrained transformers in 3 lines of code.

The simplest way to classify audio in Python.

---

Powered by pretrained transformer models (AST, Wav2Vec2, HuBERT) via HuggingFace — fine-tune a state-of-the-art audio classifier on your own dataset in a few lines of code.

from autowave import AudioClassifier

# 1. Load and train
model = AudioClassifier()
model.fit("data/train/")

# 2. Predict
result = model.predict("test.wav")
print(result)  # {"label": "dog_bark", "confidence": 0.94}

# 3. Evaluate
metrics = model.evaluate("data/test/")
print(f"Accuracy: {metrics['accuracy']:.2%}")

# 4. Save & reload
model.save("my_model/")
loaded = AudioClassifier.load("my_model/")

---

Installation

pip install AutoWave

Requirements: Python ≥ 3.10, PyTorch ≥ 2.0

---

Quick Start

1. Prepare your dataset

Organize audio files into class subfolders:

data/
  train/
    dog/     bark1.wav  bark2.wav  ...
    cat/     meow1.wav  meow2.wav  ...
    bird/    chirp1.wav chirp2.wav ...
  test/
    dog/     ...
    cat/     ...

2. Train and predict

from autowave import AudioClassifier

model = AudioClassifier()
model.fit("data/train/")
model.predict("data/test/dog/bark_test.wav")
# → {"label": "dog", "confidence": 0.97}

3. Evaluate

results = model.evaluate("data/test/")
print(f"Accuracy: {results['accuracy']:.2%}")
print(results["report"])

4. Save and reload

model.save("my_model/")
loaded = AudioClassifier.load("my_model/")
loaded.predict("new_audio.wav")

---

Zero-Shot Classification (no training)

Classify audio against any text labels — no dataset or fine-tuning required:

from autowave import ZeroShotClassifier

clf = ZeroShotClassifier()
clf.predict("audio.wav", labels=["dog barking", "cat meowing", "rain", "music"])
# → [{"label": "dog barking", "confidence": 0.91}, ...]

---

Advanced Options

model = AudioClassifier(
    model_name="ast",          # "ast" | "wav2vec2" | "hubert" | "wavlm" | any HF model ID
    epochs=10,
    batch_size=8,
    learning_rate=1e-4,
    augment=True,              # noise, pitch shift, time stretch, shift
    device="auto",             # "auto" | "cuda" | "mps" | "cpu"
    output_dir="checkpoints/",
    max_duration_s=10.0,
)
model.fit("data/train/", val_folder="data/val/")

Available models

Short name	HuggingFace model	Best for
ast (default)	MIT/ast-finetuned-audioset-10-10-0.4593	All audio types
wav2vec2	facebook/wav2vec2-base	Speech tasks
hubert	facebook/hubert-base-ls960	Speech tasks
wavlm	microsoft/wavlm-base	Speech benchmarks

Any HuggingFace AutoModelForAudioClassification-compatible model ID also works.

---

Export to ONNX

model.export_onnx("model.onnx")

---

Visualization

from autowave.visualization import plots

plots.waveform("audio.wav")
plots.spectrogram("audio.wav")
plots.mfcc("audio.wav")
plots.spectral_centroid("audio.wav")
plots.time_freq_overview("audio.wav")

---

Audio Utilities

from autowave.utils.audio import read_properties, resample, convert_format

# Metadata
props = read_properties("audio.wav")
print(props.sample_rate, props.duration_s, props.channels)

# Resample to 16 kHz
resample("audio.mp3", target_sr=16000, output_path="audio_16k.wav")

# Convert format
convert_format("audio.wav", output_format="mp3")

---

Supported Audio Formats

.wav · .mp3 · .flac · .ogg · .m4a · .aiff

---

Core Contributors

Nilesh Verma

Satyajit Pattnaik

Kalash Jindal

---

Citation

If you use AutoWave in your research or project, please cite:

@software{autowave2024,
  author       = {Verma, Nilesh and Pattnaik, Satyajit and Jindal, Kalash},
  title        = {{AutoWave}: Automatic Audio Classification with Pretrained Transformers},
  year         = {2024},
  version      = {2.0.0},
  url          = {https://github.com/TechyNilesh/Autowave},
  note         = {Python library for audio classification using AST, Wav2Vec2, HuBERT, and WavLM}
}

---

Developed with Love ❤️

Developed for ML researchers, data scientists, Python developers, speech engineers, and the open-source audio community.