THE FORECAST DESK
AI TOOLS, WATCHED DAILY AT 02:00 UTC · YOUR OVERNIGHT · GRADED IN PUBLIC

Coqui AI shut down — what happened to Coqui STT, TTS and DeepSpeech

coqui.ai · TRACK: AI VOICE CLONING · CASE CLOSED

Coqui AI shut down: the company ceased operations in 2023-12 and the founder's farewell — "Coqui is shutting down" — was posted on 2024-01-03 [FDR]. The three code lines people search for did not stop on that date, or on each other's. Coqui STT, the speech-to-text toolkit, shipped its last release, v1.4.0, on 2022-09-03 — fifteen months before the company closed. Its repository is still public and is not archived, but the README now states that the project is no longer actively maintained and that the online Model Zoo is no longer hosted [3P]. Mozilla DeepSpeech — the speech-to-text project the same founding team worked on before Coqui, per the founder's own account [FDR] — outlasted Coqui on paper and was archived read-only on 2025-06-19, its README reading "This project is now discontinued"; its last release, v0.9.3, is dated 2020-12-10 [3P]. Coqui TTS is the one line still moving: the original repository is unmaintained, but a community fork, idiap/coqui-ai-TTS, carries it forward and ships as the PyPI package coqui-tts [3P].

Coqui AI (coqui.ai) was a commercial AI voice-synthesis and voice-cloning company built by four ex-Mozilla Machine Learning Group engineers, spun out as an independent company in 2021 [FDR][3P]. It kept its core TTS stack fully open source (45.8k GitHub stars, MPL-2.0, 1,100+ languages) [3P] while selling a hosted paid layer on top [OBS]. The books close it at 2023-12 after 30 months [OBS]; the founder's farewell — "Coqui is shutting down" — was posted to the project's own discussion board on 2024-01-03 [FDR]. Cause filed: open source mistaken for a moat — the free core undercut the paid layer.

SHUT DOWN — 2023-12
lived 30 months (2021-06≈ → 2023-12) · case file below, every claim sourced

Timeline

EVIDENCE GRADE: A — primary sources on file.

Cause of death — our read

Cause filed on the roster: open source mistaken for a moat. The contradiction fits on one ledger line. The flagship repository gave away the entire capability — 1,100+ languages of TTS under MPL-2.0, now 45.8k stars [3P] — while the paid layer sold hosted access to the same thing. XTTS v1 weights were released openly; v2 followed and was, by the founder's own account, "even better" [FDR]. Each open release raised the community's floor and lowered the paid ceiling. The war chest never matched the giveaway: one round, $3.3M, on the books via 2023-03 coverage [3P], in a compute-heavy track where closed-weight rivals monetized scarcity instead of publishing it [INF]. The filed target buyers — game studios, audio post shops [3P] — were exactly the customers technical enough to self-host the free core; whoever still paid was paying for convenience, not capability, and convenience margins do not fund frontier model training [INF]. The sequence at the end is compact: final repo release 2023-12-12 [3P], death month 2023-12 after 30 months alive [OBS], founder farewell posted 2024-01-03 [FDR]. The stars outlived the company. The asset that made Coqui famous was the one it could never invoice [INF].

The lesson on file

Stars are logged as reach, not revenue. When the free artifact is the entire product — weights, code, permissive license — the paid tier must hold something the download cannot: uptime nobody wants to own, latency guarantees, compliance paper, or distribution locked to the host. Coqui held none of these; a single $3.3M round bought 30 months [OBS]. Before shipping an open core, write down the invoice line a self-hosting customer still cannot escape. If that line is blank, the open release is marketing for competitors, and the graveyard entry is pre-filed. Charge for the pain; give away the pride.

The reopen clause — what would change this verdict

Does 'open source mistaken for a moat' hold as the filed cause?

NEXT SIGNAL: A successor voice company winning on the same open weights with a different business wrapper would sharpen the lesson; primary evidence of a different proximate cause would amend the filing.

A READ WITHOUT A NEXT SIGNAL IS JUST AN OPINION — THIS ONE IS FILED IN ADVANCE, LIKE EVERYTHING ELSE ON THE DESK.

What to use instead

The AI dubbing / voice-cloning track carries a single filed entry — r008 itself — so no live substitute is on the books. Nearest tracked audio-adjacent track is AI short-video captioning, where Submagic (r007), OpusClip (r016) and Quso (r017) remain live and are filed as successes. On the code rather than the company: the TTS line continues as the community fork idiap/coqui-ai-TTS, which describes itself as a fork of the original, unmaintained repository and ships as the PyPI package coqui-tts [3P]. The speech-to-text line has no equivalent — coqui-ai/STT is unarchived but unmaintained by its own README, last released 2022-09-03, and Mozilla DeepSpeech behind it was archived on 2025-06-19 [3P]. Both STT repositories are readable and installable; neither is receiving fixes. The desk files no live speech-to-text substitute because it tracks none — this track's roster is TTS and voice cloning.

Sources on this page

Dispute a number? /corrections/ — handled in public within 48 hours.

ALL CASE FILES: WHY AI TOOLS DIE →

THE DESK COUNTS VISITS (GOOGLE ANALYTICS) ONLY IF YOU SAY YES. NO ADS, NO RESALE. /PRIVACY/