# openbst: a C reimplementation of BeSTspeech

> A from-scratch C reimplementation of the Berkeley Speech Technologies formant synthesizer sold as BeSTspeech and shipped in HumanWare Keynote GOLD. The original binaries are the comparison input and are not in the repository; make selftest checks tests/golden.txt, and make test compares the engine with those binaries.

- Type: Case study
- Targets: Binary
- Category: Independent reconstruction cases
- Relationship: Independent reconstruction project
- Sources reviewed: 2026-11-03
- Added to the catalog: 2026-11-09
- Original resource: <https://github.com/Mudb0y/openbst>
- Primary source: <https://github.com/Mudb0y/openbst/blob/a151e78f66009bc74367974d14476061f3552346/README.md>
- Catalog page: <https://awesomerea.org/resources/openbst-bestspeech-case/>
- Record JSON: <https://awesomerea.org/resources/openbst-bestspeech-case/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/resources/openbst-bestspeech-case/index.md)

## Requirements & limits

- **Prerequisites:** Commit a151e78f66009bc74367974d14476061f3552346. The README names twenty builds: the 1995 32-bit English engine, six 1998 16-bit modules (English, Dutch, French, German, Italian, and Spanish), and thirteen 2006 32-bit builds. Put the original binaries under dll/1995, dll/1998, and dll/2006. They are not in the repository. make still writes build/libbst.a, build/libbst.so, and build/bstspeak, because src/data holds the lifted tables.
- **Testing evidence:** The README says the 1995 build is bit-exact on 24 utterances, the six 1998 modules match 226 words and 20793 frames with 144 utterances sample for sample, the thirteen 2006 builds match 2358 utterances, and tests/lifttest.sh checks 411894 utterances. Those runs need the original binaries. make test is that gate. Without the binaries, make selftest checks tests/golden.txt, which the README says holds sample counts and hashes for 4906 utterances written from the binaries by make golden.
- **License:** LICENSE puts the code, tests, and tools under the MIT license, copyright 2026 Stanislaw Przedzinkowski, and says the tables read from Berkeley Speech Technologies’ binaries, later HumanWare’s, are not covered by that license.

## Primary sources

1. <https://github.com/Mudb0y/openbst/blob/a151e78f66009bc74367974d14476061f3552346/README.md>
2. <https://github.com/Mudb0y/openbst/blob/a151e78f66009bc74367974d14476061f3552346/LICENSE>

## More in Native binaries

- [OpenTV: a portable C rebuild of Centigram TruVoice](https://awesomerea.org/resources/opentv-truvoice-case/index.md) · Case study
- [L&H TruVoice: a portable C rebuild of tv_enua.dll](https://awesomerea.org/resources/lh-truvoice-decomp-case/index.md) · Case study
- [Microsoft Anna: a bit-exact C rebuild of the TTS20 engine](https://awesomerea.org/resources/ms-anna-tts-case/index.md) · Case study
- [OneCore TTS: a C rebuild of David, Zira and Mark](https://awesomerea.org/resources/onecore-david-zira-mark-case/index.md) · Case study
- [Microsoft Sam: a portable C rebuild of the SAPI 5 engine](https://awesomerea.org/resources/ms-sam-mike-mary-case/index.md) · Case study
- [Eloquence 6.1: a C rebuild of ECI.DLL](https://awesomerea.org/resources/eloquence-61-eci-case/index.md) · Case study

## About this entry

The description was reviewed against primary sources.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.