108 non-neural speech synthesis works, 1979–2026: from vocal-tract physics to WORLD and STRAIGHT

Asked (summary):

What peer-reviewed papers, open-source systems, and technical reports through October 2026 describe non-neural text-to-speech synthesis — articulatory synthesis, formant synthesis, concatenative/unit-selection synthesis, source-filter models, and non-neural vocoders such as WORLD and STRAIGHT? For each, give title, authors, year, venue, method, reported naturalness or intelligibility findings, and limitations, as a table with a source URL.

The scan recovered 108 substantive works dated 1979–2026, each with a stated evaluation finding and a stated limitation, drawn from 49 independent source hosts (no single host supplies more than 11.1% of rows, no single URL more than 2.8%). Coverage is broad but not mathematically exhaustive: only works for which a reported quality, intelligibility, or naturalness finding plus a limitation could be recovered are included, and historical terminology overlaps.

Every work, by year and method family

Each square is one work; hover or tap for title, venue, and finding; click a square to jump to its row. Colour marks the method family; lane labels carry counts. Peer-reviewed papers, technical reports, and project documentation are distinguished by the venue text in the table.

Articulatory and physical-model synthesis is the longest-running thread (24 works): a 1983 Florida dissertation needed about five hours of computation per second of speech ufdc.ufl.edu, and by 2021 VocalTractLab syntheses were rated intelligible yet still behind commercial systems arxiv.org.

The 2010s hold 56 of the 108 works — the peak of HMM statistical-parametric synthesis and vocoder research, a period in which the WORLD vocoder became a standard high-continuity baseline, though it was later outperformed in quality by neural WaveNet vocoders arxiv.org.

Unit-selection concatenative synthesis repeatedly won naturalness comparisons against rule-driven formant systems, at the cost of large corpora and audible join errors — a trade-off cited as a limitation across the 24 concatenative works in the table.

The most recent entry, DYNARTmo (arXiv, October 2026), is a preprint with the author's informal listening assessment only — not peer-reviewed or independently evaluated arxiv.org.

The complete bibliography

Title	Authors	Year	Venue	Method	Findings	Limitations	Source URL

Authors are blank where the retrieved primary or official page did not expose them. Long method, findings and limitations texts are collapsed — tap “…more” to expand. The year and family filters above also drive the chart.

Method: bibliography of 108 non-neural speech-synthesis works (peer-reviewed papers, technical reports, and open-source project documentation), 1979–2026, compiled by a research scan completed through October 2026 and drawn from 49 independent hosts; max one-host share 11.1%, max one-URL share 2.8%. Each row carries a reported quality, intelligibility, or naturalness finding and a stated limitation taken from its source; works without a recoverable finding plus limitation were excluded. Method-family labels were normalized into five groups for the chart; original family wording varies. Some long texts are truncated at source.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.