Coqui

» see coqui.ai as it lived, on the wayback machine

~ the last words ~

Mozilla shut its machine learning group in August 2020, which ended work on DeepSpeech and on the text-to-speech project that sat next to it. Four of the people who had been doing that work, Eren Gölge, Josh Meyer, Kelly Davis and Reuben Morais, announced Coqui on the Mozilla forum on 15 March 2021 and went to Berlin to carry on building the same thing with a company around it.

What they built got very good very fast. The Coqui TTS library became the standard answer to "how do I generate speech on my own hardware" for everybody unwilling to post their audio to somebody else's API, and XTTS, the model they released in 2023, could clone a voice from a few seconds of a recording and then speak seventeen languages in it. The seed round was $3.3 million, led by ScaleX Ventures out of Istanbul with Mango Capital and DNX Ventures alongside.

The commercial side was Coqui Studio, a browser tool for voice-over work, plus a REST API. On 25 December 2023 the homepage was still selling it: pricing starts at $1 per day, five named synthetic performers listed with their characters, a line at the bottom reading "Made with love in Berlin". Nine days later it was a single sentence.

~ the record ~

~ one sentence on the third of January ~

The Archive dates this to the hour. At 25 December 2023 the page is 17,819 bytes of product. At 3 January 2024 06:35 UTC it is 10,590 bytes, and the entire body reads: "Coqui is shutting down. Thank you for all your support!" with a heart after it. The footer underneath still says Made with love in Berlin, still carries the 2023 copyright line, and still links the community forum.

Nobody wrote a long letter. Gölge posted the same sentence to LinkedIn. When a user asked in the GitHub discussions six weeks later whether anything was salvageable, a maintainer answered in eleven words: "Coqui is shut down yes. the future of the project is unknown."

~ the istanbul line on the cap table ~

This one lands closer to home than most things I write up here, and the connection is on paper rather than in my memory. Gölge did his degrees at Bilkent, in the city I work in now, and the fund that led the seed sits in Istanbul. Edge inference is my working problem, and a voice model that fits on hardware you already own instead of hardware you rent by the hour is exactly the artifact I wanted this company to go on producing.

It stayed a paper connection. Coqui was a Berlin company selling to American studios and game developers, and I never had a Coqui Studio account or a reason for one. What I have instead is the library, which turned up in my working life the way it turned up in everyone's, as a pip install with no company attached to it.

~ what an open model could not charge for ~

Christopher Manning posted the epitaph the same week, and he was not being unkind about it: the lesson is probably that it is just really hard to succeed as a standalone text-to-speech company. The model that makes you famous is the model anyone can download, and the customers who would pay for the hosted version are the ones with enough engineering not to need it.

The repository is archived and the weights are still on Hugging Face, where they get pulled more now than they did while there was a company to pay for the bandwidth. Forks carry the maintenance.

coqui.ai answers today with the GitHub Pages 404, "Site not found", because the DNS still points at a static host with nothing published behind it. No registrar took the name and no squatter wanted it. The team simply stopped deploying.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »