Research map / Architectures and efficiency / RNNs and normalization
Efficient Neural Audio Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the 60-layer WaveNet with a single-layer WaveRNN-896 using dual softmax | WaveRNN · Text-to-speech held-out test set | A/B comparison overall score: 0.02 ± 0.08 | No clear effect |
| Increase hidden-state size from 896 to 2048 | WaveRNN · Text-to-speech held-out test set | NLL: 5.42 -> 5.33 | Helped |
| Use a large sparse model rather than a small dense model at the same parameter count | WaveRNN · Text-to-speech held-out test set | NLL | Helped |
| Use 16×1 rather than 4×4 structured sparsity (ablation) | Sparse WaveRNN · Text-to-speech sampling on mobile CPUs | samples/sec: 31,400 vs 29,100; 21,600 vs 19,800 | Helped |
| Generate 16 samples per step with subscaling | Subscale WaveRNN · Text-to-speech held-out test set | A/B comparison overall score: -0.03 ± 0.05 | No clear effect |
| Increase subscaling batch factor from 8× to 16× (ablation) | Subscale WaveRNN · Text-to-speech held-out test set | MOS: 4.39 ± 0.06 -> 4.30 ± 0.08 | Hurt |
| Fuse subscaling into the WaveRNN to produce 32 bits per step (ablation) | Fused Subscale WaveRNN · Text-to-speech sampling on Nvidia P100 | real-time speed: 4× -> 10× | Helped |
| Fuse subscaling by splitting the hidden state (ablation) | Fused Subscale WaveRNN · Text-to-speech held-out test set | MOS: 4.37 ± 0.07 -> 4.31 ± 0.08 | Hurt |
| Use a single softmax over all 2^16 sample values instead of dual softmax (ablation) | WaveRNN · Text-to-speech held-out test set | NLL: 1-2 centi-nats worse | Hurt |
| Condition the hidden state only on the previous input and use a separate layer to condition fine bits on coarse bits (ablation) | WaveRNN · Text-to-speech held-out test set | NLL: 1-2 centi-nats worse | Hurt |
| Increase subscaling batch factor beyond 2× in the fused architecture (ablation) | Fused Subscale WaveRNN · Text-to-speech held-out test set | audio quality | Hurt |
Closest papers by meaning
- SampleRNN: An Unconditional End-to-End Neural Audio Generation Model (ICLR 2017 · similarity 0.91)
- Parallel WaveNet: Fast High-Fidelity Speech Synthesis (ICML 2018 · similarity 0.89)
- SING: Symbol-to-Instrument Neural Generator (NeurIPS 2018 · similarity 0.89)
- FloWaveNet : A Generative Flow for Raw Audio (ICML 2019 · similarity 0.88)
- Deep Voice: Real-time Neural Text-to-Speech (ICML 2017 · similarity 0.88)
- PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.