Research map / Image generation / Autoregressive and latent generation
Gogo: Group-wise granularity-ordered codec for stable and efficient speech generation
Weidong Chen, Helen M. Meng, Xixin Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 group-wise quantization instead of frame-wise RVQ (token position 1) | Gogo tokens · AR perplexity | perplexity: 2.3 -> 0.9 | Improved |
| use group-wise quantization instead of frame-wise RVQ (token position 10) | Gogo tokens · AR perplexity | perplexity: 691.4 -> 228.3 | Improved |
| use group-wise instead of single-layer frame-wise VQ for the coarsest token | Gogo tokens · AR perplexity | perplexity: 247.8 -> 0.9 | Improved |
| add GRPO token allocator to GogoSpeech (47 -> 36 tokens/s) (ablation) | GogoSpeech · Seed-TTS test-en | RTF: 0.535 -> 0.455 | Improved |
| add GRPO token allocator to GogoSpeech (ablation) | GogoSpeech · Seed-TTS test-en | WER: 2.394 -> 2.469 | Hurt |
| add GRPO token allocator to GogoSpeech (ablation) | GogoSpeech · Seed-TTS test-en | SMOS: 4.381 -> 4.253 | Hurt |
| add GRPO token allocator to GogoSpeech (ablation) | GogoSpeech · Seed-TTS test-en long-form | WER: 1.788 -> 1.845 | Hurt |
| use two-stage coarse-to-fine generation (GogoSpeech) versus F5-TTS on long-form speech | GogoSpeech · Seed-TTS test-en long-form | WER: 1.812 -> 1.788 | Improved |
| use Gogo at 47 TPS vs X-codec2 at 50 TPS | Gogo · LibriTTS test-clean | WER: 6.40 -> 6.35 | Improved |
| retain more than six tokens per group at reconstruction | Gogo · LibriTTS test-clean | WER | No clear effect |
| retain more than four tokens per group at reconstruction | Gogo · LibriTTS test-clean | PESQ | Improved |
| move probing from coarse token 1 to fine token 10 for pitch | Gogo tokens · LibriTTS probing | probe MSE (pitch): 1455 -> 1012 | Improved |
Closest papers by meaning
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.85)
- CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech (ICLR 2024 · similarity 0.85)
- ELLA-V: Stable Neural Codec Language Modeling with Alignment-Guided Sequence Reordering (AAAI 2025 · similarity 0.85)
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 2025 · similarity 0.84)
- FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks (NeurIPS 2025 · similarity 0.84)
- SECodec: Structural Entropy-based Compressive Speech Representation Codec for Speech Language Models (AAAI 2025 · similarity 0.84)
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.
