Reproduce “DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Roi Benita, Michael Elad, Joseph Keshet
Researchers at Technion – Israel Institute of Technology introduced DiffAR, an end-to-end denoising diffusion autoregressive model for raw speech waveform generation. This model synthesizes speech of unlimited duration by generating overlapping frames and effectively reproduces fine-grained acoustic details like vocal fry, outperforming existing models in subjective quality and intelligibility on the LJ-Speech dataset.
Key insights
- Direct raw waveform generation via diffusion models can effectively capture and reproduce complex, fine-grained acoustic phenomena like vocal fry, which are often missed by spectrogram-based approaches.
- An autoregressive frame-based diffusion strategy is effective for synthesizing continuous, arbitrary-length speech while maintaining high temporal coherence and quality.
- Diffusion models offer inherent stochasticity, allowing for diverse and natural-sounding speech realizations from the same input, which can be balanced with explicit linguistic conditioning for control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.