eigenwave-asr 1.0.4

EigenWave-ASR: High-Performance Speech Recognition with Multi-Scale Robin Features

EigenWave-ASR 🎤

High-Performance Speech Recognition with Multi-Scale Robin Features

A novel ASR model achieving 6.36% WER on LibriSpeech test-clean with only 27.8M parameters.

🚀 Quick Start

from eigenwave import EigenWaveASR

# Load model
model = EigenWaveASR.from_pretrained("./")  # local directory
# or
model = EigenWaveASR.from_pretrained("sakibhasan/eigenwave-asr")  # from Hub

# Transcribe audio
text = model.transcribe("speech.wav")
print(text)
# Output: "hello this is a test of the speech recognition system"

📦 Installation

pip install eigenwave-asr

# For best accuracy (KenLM language model):
pip install eigenwave-asr[lm]

# For HuggingFace Hub support:
pip install eigenwave-asr[all]

📋 Usage Examples

Basic Transcription

from eigenwave import EigenWaveASR

model = EigenWaveASR.from_pretrained("sakibhasan/eigenwave-asr")

# From file
text = model.transcribe("audio.wav")

# Batch
texts = model.transcribe(["audio1.wav", "audio2.wav", "audio3.wav"])

# From tensor (16kHz mono)
import torch
audio_tensor = torch.randn(16000 * 5)  # 5 seconds
text = model.transcribe(audio_tensor)

Detailed Output

result = model.transcribe_with_details("audio.wav")
print(result)
# {
#     "text": "hello world",
#     "duration": 2.5,
#     "processing_time": 0.320,
#     "rtf": 0.128,         # Real-Time Factor (< 1.0 = real-time)
#     "real_time": True
# }

CPU Inference

model = EigenWaveASR.from_pretrained("sakibhasan/eigenwave-asr", device="cpu")
text = model.transcribe("audio.wav", beam_width=10)  # smaller beam for speed

Without Language Model (faster, less accurate)

model = EigenWaveASR.from_pretrained("sakibhasan/eigenwave-asr", use_lm=False)
text = model.transcribe("audio.wav")

📊 Performance

Dataset	Greedy WER	Beam+LM WER
test-clean	~8.5%	6.36%
test-other	~20%	~15%

🏗️ Model Details

• Architecture: Conformer-style encoder with Multi-Scale Robin Features
• Parameters: 27.8M
• Input: 16kHz mono audio
• Output: English text (lowercase)
• Training: LibriSpeech 960h, 182k steps
• Features: Novel Robin differential operator feature extraction at scales [1, 3, 5]

⚡ Optimal Hyperparameters (Optuna-tuned)

alpha       = 0.9268   (LM weight)
beta        = 0.061    (word insertion bonus)
temperature = 0.536    (softmax sharpness)
beam_width  = 50       (beam search width)

📄 License

MIT License

👤 Author

Sakib Hasan (@sakibhasanml)