# Search scope and evidence notes Checked: 26 October 2026. Language of the deliverable: English. Intended audience: a professor discussing research directions. ## Coverage The catalog contains 59 primary-source records: 32 on speech synthesis and 27 on speech deepfake detection or adjacent provenance questions. The first-publication window is 2025–2026. Latest checked arXiv revisions are pinned in links. Proceedings dates are used for two official venue records. This is a broad, question-led literature collection. It is not a preregistered systematic review, a complete index of every conference paper, or a universal SOTA leaderboard. Coverage cannot establish that no relevant work was missed. Commercial systems without a verifiable technical paper are not treated as literature records. ## Discovery and verification - Primary sources: arXiv records and export API; ACL Anthology; ISCA Archive; selected official OpenReview documents, author project pages and model releases. - Search themes: zero-shot TTS; independent voice/style/emotion control; speech tokenization and content representation; joint training and preference optimization; streaming text input; generation latency; long-form synthesis and evaluation. - Detection themes: unseen generators and speakers; self-supervised speech backbones; continual learning; compact detectors; partial neural edits; temporal localization; evidence attribution; multilingual transfer; compression/replay; expressive attacks; mixed audio; generator adaptation; watermark interactions. - Representative discovery queries included `speech synthesis October 2026 text-to-speech benchmark reinforcement`, `audio deepfake detection October 2026 unseen generator localization`, `speech tokenizer 2026`, and exact paper/model names. - Additional exact-title checks included ASVspoof 5, AUDDT, AT-ADD, AffectDF, GenTraceBench, VoXtream2, Luna-TTS, LoCA, FireRedAudio, PartialEdit, XY-Tokenizer, and Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios. - The source-metadata XML caches the arXiv API response for 57 requested records. Two further records use official conference pages. Titles, author lists, publication dates and version identifiers came from those primary records. Search engines also returned third-party summaries. These were discovery leads, not evidence for technical claims. All retained catalog entries link to primary sources. ## Selection and grouping Included: synthesis systems, relevant representation studies, training methods, delivery methods, evaluation resources, detection methods, robustness studies and two adjacent provenance studies. Voice conversion is included where it helps define synthesis or detector threats. Excluded as primary topics: full-duplex conversational agents, ASR-only systems, music-only generation, visual deepfakes and product announcements without a checked paper. StreamAlign is included as adjacent streaming representation work, with its speech-continuation scope stated. AT-ADD's evaluation-plan and challenge-summary papers are represented by its expanded benchmark report, not counted as three independent contributions. ASVspoof 5 has separate resource and evaluation reports, explicitly identified as one benchmark family. The Evolving Bottleneck is marked as background analysis rather than a new empirical method. Spark-TTS is a 2025 historical anchor. ## What was verified All 59 entries: primary bibliographic record, title, author metadata, public date, source URL, and abstract-level problem and method. The catalog uses short paraphrases, not copied abstracts. Selected full-text checks support caveats around Kanade's streaming scope, VoxCPM2's representation, staged end-to-end TTS training, Qwen-Audio-3.0-TTS training, Luna-TTS latency protocol, Drama 3 evaluation retries, and detector cross-corpus behavior. Most of the broader catalog remains abstract-level review. No experiments were independently reproduced. Venue claims are deliberately sparse; an arXiv link does not imply peer review. The `Evidence boundary` line combines reported scope with editorial advice about interpretation. Suggested experiments and research opportunities are the map author's proposals. They are not claims that the cited authors ran those experiments. ## Corrections made while checking - AUDDT's checked v2 title is **A Unified Benchmark Toolkit for Audio and Speech Deepfake Detectors**, and its abstract covers 31 datasets. - The checked CAFNet paper is **Half-Truth Audio Detection and Localisation: A Lightweight Cross-Attentive Architecture and a Cross-Corpus Diagnostic Study**, v3. Its cross-corpus results are mixed; binary transfer to PartialSpoof is near chance. - FireRedAudio's verified identifier is **2608.24168**. Other similar-looking identifiers are not substituted. - VoXtream's first submission is in 2025, with a 2026 revision; these dates are distinguished. - Kanade was first submitted on 31 January 2026 despite its 2602 arXiv identifier. - A lower token rate is not automatically a lower bitrate or a shorter end-to-end response delay. - A recovered watermark, blind detector score, and proof of authenticity are different kinds of evidence. ## Limitations and follow-up The catalog emphasizes the requested themes and visible 2026 work. It does not audit training-data overlap or reproduce model rankings. Positive abstract claims may not survive a stronger matched-budget evaluation. Patent literature, non-public industrial systems and every regional venue are outside the collection. Before selecting a thesis problem, read the full methods and limitations of a small target subset, confirm artifact availability and licenses, and reproduce one baseline under a split that isolates the proposed failure.