A living systematic review and knowledge graph of the state of the art in text-to-speech (TTS), voice conversion (VC), and spoken conversational agents (SCA). Papers are ingested on a rolling basis, enabling both current-state snapshots and year-on-year trend analysis.
Venues: Interspeech, ICASSP, ACL, EMNLP, NAACL, NeurIPS, ICLR, ICML, ASRU, SLT, arXiv preprints, and technical reports from industry labs (Google, Microsoft, Meta, ElevenLabs, Apple, Amazon, and others).
Period: August 2025 onward, with foundational papers added via citation discovery.
Corpus: ~800 accepted papers; pages added continuously as ingestion proceeds.
index.md Landing page — concept navigation, links to all sections
overview.md Evolving synthesis of dominant paradigms and emerging trends
log.md Reverse-chronological log of ingests, integrations, and queries
papers/ One page per ingested paper — method, results, claims, novelty assessment
index.md Full paper catalog
concepts/ Technology and method concept pages
index.md Concept directory
_evidence/ Machine-oriented evidence digests (one per concept; used for synthesis)
comparisons/ Cross-paper comparison tables generated in response to research queries
venues/ Per-venue summary pages (named {year}-{venue}, e.g. 2025-interspeech)
index.md Venue directory
reports/ Periodic field reports — monthly, quarterly, yearly
| Area | Concepts |
|---|---|
| Core architectures | Flow matching · Diffusion · Autoregressive codec TTS · Transformer enc-dec · GAN vocoder |
| Capabilities | Zero-shot TTS · Voice conversion · Multilingual TTS · Emotion synthesis · Prosody control · Streaming TTS · Instruction-conditioned TTS |
| Systems | Spoken language model · Speech-to-speech |
| Foundations | Neural codec · Self-supervised speech · Disentanglement · Speaker adaptation · RLHF for speech |
| Evaluation | Evaluation metrics · Subjective evaluation |
Each concept page includes: executive summary, current status, methods and variants, major claims (strongly supported / emerging / contested), relationship to other concepts, representative papers, and a trend summary.
Each paper page includes:
- Paper card — venue, year, authors, paper link, and one-sentence contribution in a single callout
- Method — system description with embedded architecture figure where available
- Claims — 2–5 generalised propositions about the field that this paper supports, weakens, or complicates
- Field significance — level (low / moderate / high / foundational) and contribution type
- Novelty assessment — honest evaluation of what is genuinely new vs. incremental
- Limitations and open questions
This repo is the output of an automated ingestion pipeline. Sources, scripts, metadata, and agent definitions live in the companion infra repo: speech-generation-wiki-infra.
Each paper page is generated by an LLM ingest agent that reads the full parsed PDF, writes a structured wiki page, and selectively embeds architecture diagrams. A separate integration agent updates concept pages, cross-links citing/cited paper pairs, and maintains concept evidence digests. All metric values are sourced directly from paper tables — nothing is estimated.