Offline TTS in 2026?

Comments

• fhein@lemmy.world ⁨11⁩ ⁨hours⁩ ago

  https://github.com/resemble-ai/chatterbox is pretty good, and has both TTS and voice cloning. Main disadvantage for me was that even if the cloning gives a consistent voice, the generated samples can get random accents.

  https://huggingface.co/zai-org/GLM-TTS also seemed pretty promising, but I haven’t had time to test it yet.

  source
• IcedRaktajino@startrek.website ⁨2⁩ ⁨days⁩ ago

  github.com/marytts/marytts

  I’ve used MaryTTS semi-recently. It’s older but works well enough for my cases. I have it running on a server (locally) and my endpoints make a call to it and playback the returned audio file.

  On Android, I use SherpaTTS which has good voices, but I’m not aware of a desktop/Linux option. It mentions using voices from Coqui which you linked, so I would guess that would be the way to go for desktop.

  source

  • comrade_twisty@feddit.org ⁨2⁩ ⁨days⁩ ago

    SherpaTTS is great on GrapheneOS with OSM for navigation.

    source

    • warmaster@lemmy.world ⁨1⁩ ⁨day⁩ ago

      I’m using it with CoMaps, freaking great.

      source
    • FrostyPolicy@suppo.fi ⁨2⁩ ⁨days⁩ ago

      +1 for this.

      source
  • otter@lemmy.ca ⁨2⁩ ⁨days⁩ ago

    Sherpa links to this page, if anyone wants to preview what the voices sound like

    huggingface.co/spaces/k2-fsa/text-to-speech

    source
• ikidd@lemmy.world ⁨1⁩ ⁨day⁩ ago

  Piper just moved to github.com/OHF-Voice/piper1-gpl

  It’s fine, and it’s probably the best TTS you’re going to run locally.

  source
• solrize@lemmy.ml ⁨2⁩ ⁨days⁩ ago

  Tbh I generally prefer that robot voices actually sound like robots. I don’t want human sounding TTS since I don’t like anthropomorphized machines. So I’d stay with something low tech.

  source

  • Buckshot@programming.dev ⁨1⁩ ⁨day⁩ ago

    100%. It irks when software has messages in the first person. The human-like TTS are very uncanny valley to me.

    I needed voice over for a video recently and everyone is telling me to use AI, I tried dozens and they’ll all just off. Ended up getting a friend to record it. I don’t have a decent mic so my own recording sounded bad.

    source

    • Teppichbrand@feddit.org ⁨1⁩ ⁨day⁩ ago

      I went back from ShepherdTTS to eSpeak because I liked the robot woman better. She’s easy to understand, takes up 5 mb on my phone and it’s a fun conversation starter.

      source
• Samsy@lemmy.ml ⁨2⁩ ⁨days⁩ ago

  SherpaTTS?

  f-droid.org/packages/org.woheller69.ttsengine

  Got regular updates.

  source
• muzzle@lemmy.zip ⁨2⁩ ⁨days⁩ ago

  for android: SherpaTTS or whisper

  source
• damnthefilibuster@lemmy.world ⁨2⁩ ⁨days⁩ ago

  Kokoro is your best bet right now. It works wonderfully even in a docker container with no GPU. There are others but I don’t have the list right now. Will throw another update on here when I do.

  The rhasspy guy was very invested in Coqui. He built a lot of his own stuff, for his home automation and such. But Coqui was superior, so he started spending time on that.

  Unfortunately, the coqui team (based out of Mozilla) was very distracted and didn’t ship a lot of stuff on time or at all. It doesn’t even have basic stuff like SSML support right now, if I recall correctly. So the rhasspy guy also lost steam.

  Of course, with the OpenAI model of audio generation, you’re expected to not use SSML at all and just use the black box API to get “good enough” results. That really sucks.

  Oh, I just remembered which other one I wanted to mention - someone has built an open source version of NotebookLLM, complete with multi voice support. But it requires GPU, I believe. Do what you will with that. I’ll add a link if I find it.

  I prefer kokoro because it’s really solid and works really well on CPU.

  source

  • markko@lemmy.world ⁨1⁩ ⁨day⁩ ago

    I’m guessing this is what you’re talking about?

    source

    • damnthefilibuster@lemmy.world ⁨1⁩ ⁨day⁩ ago

      I dont think so. This one doesn’t have any code or implementation details. The one I saw was fully installable but a PITA.

      source
• CubitOom@infosec.pub ⁨2⁩ ⁨days⁩ ago

  Here is the repo for the piper I use, github.com/OHF-Voice/piper1-gpl

  For my phone, I use this tts engine, k2-fsa.github.io/sherpa/onnx/tts/apk-engine.html

  source

  • damnthefilibuster@lemmy.world ⁨2⁩ ⁨days⁩ ago

    What’s your workflow on your phone?

    source

    • CubitOom@infosec.pub ⁨2⁩ ⁨days⁩ ago

      I have kaldi auto updated with my preferred voice when ever there is a new release via obtainium.

      I set kaldi as my tts engine, and I disabled the google TTS.

      Then anytime I use TTS on my phone it uses kaldi.

      source

      • -> View More Comments
• i_am_not_a_robot@discuss.tchncs.de ⁨2⁩ ⁨days⁩ ago

  Check the README for piper. It moved to github.com/OHF-Voice/piper1-gpl

  source
• handsoffmydata@lemmy.zip ⁨2⁩ ⁨days⁩ ago

  I highly recommend Mlx-audio for anyone doing tts on Apple Silicon. It offers great performance, leverages kokoro-82M, and plays well with streaming frontends like Open webui. The one shot voice cloning feature is also pretty cool.

  source

  • Rhaedas@fedia.io ⁨2⁩ ⁨days⁩ ago

    Kokoro was the one I was going to mention. I played around with it a bit, was very impressed with the speed and quality. And then I realized I had been using it in CPU mode. GPU is incredible.

    source
  • damnthefilibuster@lemmy.world ⁨2⁩ ⁨days⁩ ago

    That looks amazing! Will check it out. Love Kokoro and love how good Apple Silicon is!

    source
• Konraddo@lemmy.world ⁨2⁩ ⁨days⁩ ago

  Try alltalk_tts v2. One of the features is you can provide an audio sample and the AI will imitate the voice. The overall quality is pretty good, if you choose a larger model and let it run.

  source

  • danielquinn@lemmy.ca ⁨1⁩ ⁨day⁩ ago

    Ooh! Has anyone managed to do this with Majel Barrett’s (the Enterprise computer) voice yet?

    source
• early_riser@lemmy.world ⁨1⁩ ⁨day⁩ ago

  None of this may be relevant, but I’m curious what your use case is. I use TTS very extensively to consume media and have my preferences. None of them are open source, but as far as I know all operate locally, though they’re baked into other programs like screen readers and ebook readers.

  I prefer older more robotic voices because they remain intelligible at high speed. Eloquence is a favorite, as are the older Apple voices like Fred and Ralph. I think it has gone by other names but TruVoice (spacing and capitalization may vary) is also up there. It was semi popular during the surreal meme era. Another memetic synth that’s a little before my time but I quite enjoy is DECTalk (AKA the Moonbase Alpha voice). I believe Vocalizer was responsible for the OG Siri voice Samantha and that one’s a more human voice that’s still serviceable at high speeds.

  source
• RheumatoidArthritis@mander.xyz ⁨2⁩ ⁨days⁩ ago

  Save that post for the next time when someone with too much time on their hands asks what project they should start/contribute to.

  source
• NarrativeBear@lemmy.world ⁨2⁩ ⁨days⁩ ago

  Faster Whisper could be a option, there is various GUI options available as well.

  github.com/SYSTRAN/faster-whisper

  And if you are looking for something that you can “just install”, I recommend balabolka. The voices are natural and you can use some of the windows built in voices to make it more natural.

  www.cross-plus-a.com/balabolka.htm

  source
• eodur@piefed.social ⁨2⁩ ⁨days⁩ ago

  It really depends on what you want to do with it. I run wyoming-piper as part of my Home Assistant deployment and its been rock solid. The Wyoming protocol is pretty well documented too, so you should be able to integrate with it pretty easily.

  source
• just_another_person@lemmy.world ⁨2⁩ ⁨days⁩ ago

  Not sure what you’re asking here, but are you talking about the voice part, the TTS pat, or the interaction?

  source

  • flux@lemmy.ml ⁨2⁩ ⁨days⁩ ago

    I’m curious how could it be unclear that this post is about the TTS part? Espeak is provided as an example.

    source