Research map / Image generation / Autoregressive and latent generation
SampleRNN: An Unconditional End-to-End Neural Audio Generation Model
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 |
|---|---|---|---|
| use hierarchical SampleRNN instead of a single RNN | raw-waveform audio generation · Blizzard, Onomatopoeia, and Music test sets | NLL (bits): Blizzard: 1.434 -> 1.387; Onomatopoeia: 2.034 -> 1.990; Music: 1.410 -> 1.076 | Helped |
| use hierarchical SampleRNN instead of the re-implemented WaveNet | raw-waveform audio generation · Blizzard and Onomatopoeia test sets | NLL (bits): Blizzard: 1.480 -> 1.387; Onomatopoeia: 2.285 -> 1.990 | Helped |
| use hierarchical SampleRNN instead of the re-implemented WaveNet | raw-waveform audio generation · Music test | NLL (bits): 1.464 -> 1.076 | Helped |
| use SampleRNN-GMM instead of RNN-GMM | real-valued audio modeling · Blizzard test | Average Test NLL: -2.415 -> -2.782 | Helped |
| increase truncated-BPTT subsequence length (ablation) | SampleRNN · Blizzard validation | NLL (bits per audio sample): 1.575 -> 1.364 | Helped |
| add input embeddings for quantized samples (ablation) | SampleRNN (2-tier) · Blizzard test | NLL: 1.566 -> 1.392 | Helped |
| add input embeddings for quantized samples (ablation) | SampleRNN (2-tier) · Blizzard validation | NLL: 1.539 -> 1.369 | Helped |
| use a sample-level autoregressive module instead of conditionally independent Multi-Softmax outputs (ablation) | SampleRNN (2-tier) · Blizzard test | NLL: 1.685 -> 1.392 | Helped |
| use a sample-level autoregressive module instead of conditionally independent Multi-Softmax outputs (ablation) | SampleRNN (2-tier) · Blizzard validation | NLL: 1.656 -> 1.369 | Helped |
| replace SampleRNN with a SampleRNN-WaveNet hybrid (ablation) | raw-waveform audio generation · preliminary experiments | generation performance | Hurt |
| retain recurrent state across a one-second silence interruption | SampleRNN (3-tier) · two-speaker audiobook samples | same-speaker classification across segments: 83% | Helped |
Closest papers by meaning
- AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation (NeurIPS 2026 · similarity 0.88)
- UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities (NeurIPS 2026 · similarity 0.87)
- UniAudio: Towards Universal Audio Generation with Large Language Models (ICML 2024 · similarity 0.87)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · similarity 0.86)
- Masked Audio Generation using a Single Non-Autoregressive Transformer (ICLR 2024 · similarity 0.86)
- Fast Timing-Conditioned Latent Audio Diffusion (ICML 2024 · similarity 0.85)
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.