ICASSP 2018 Reproduction arXiv:1712.01120 Apache 2.0

WaveNet Based Low Rate Speech Coding

This is not an officially supported Google product. Interactive comparison studio for Closed-Loop WaveNet Waveform Coding (WW) and 2.4 kb/s Codec 2 Conditioned Parametric WaveNet Coding (W_wo, W_w) with implicit 4–8 kHz bandwidth extension on real LibriSpeech test utterances (run python3 hf_space/app.py locally for live Gradio encoding).

Coder Configuration

Spoken Transcript:
"The professor knew whom he had to deal with."
Decoded 16 kHz Speech (W_w) 2.40 kb/s
Reference 16 kHz Uncompressed Speech 128.00 kb/s (8-bit μ-law)

Waveform & Instantaneous Conditional Entropy Profile H_i (Eq. 2)

Reference 16 kHz Speech Decoded 16 kHz Speech WaveNet Instantaneous Entropy H_i (bits/sample)
—
Bitrate
—
Mean Entropy H_bar (Eq. 4)
—
Cross-Entropy R (Eq. 5)
—
PESQ-WB (P.862.2)
—
STOI Intelligibility
—
Composite MOS-LQO
—
Log-Spectral Distance
—
4-8 kHz High-Band Energy