Local, CPU-Friendly, High-Quality TTS (Text-to-Speech) with Kokoro | Hacker News Reader

Local, CPU-Friendly, High-Quality TTS (Text-to-Speech) with Kokoro

ariya.io

515 pointsspeckx98 comments

I have used Kokoro fairly extensively for an accessibility product. I have loved working with it (especially because I don't have an NVidia GPU like many TTS of similar quality require).

I particularly appreciate the fact that it lets you manually add IPA pronunciation guides. There have been some cases where an important word is a homograph and Kokoro assumed the wrong pronunciation.

The place where it falls a little short is in saying just a single word or two. Try having it say simply "six" and it almost always says something like "ah-six-ah". I found a way around that though. If you give it a longer sentence to say (eg "The word is: six") it will say it fine. The trick is that the Kokoro API gives you the timestamp of each word in the sentence. So you can have a Python script crop out just the word you care about. The intonation is a little flat this way, but is very reliable.

I asked about this on the discord, and was told that it is a limitation of the small parameter size. But in fairness to Kokoro, even eleven-labs' voices suffer from this occasionally.

Great technique, thanks for sharing

Same trick with many others. I say "Knight" to Wispr or Google and see "night". I say "Knight to f3", and just backspace out "to f3".

Of course, over time, I use custom entries.

My snippet expansion entry in Wispr is "Chess Knight" = "Knight" ("Knight to f3" without customization was more reliable than "Chess Knight")

I also use "dot bullet" = "•", as I like to separate thoughts with • more than ;

It is indeed a common weakness of TTS models.

Unfortunately it makes it unsuited for my use case, which is almost entirely single words, as I don't particularly want to deal with stitching/segmenting input/output.

Fun... This is something I actually care about...

I used to keep a version of whisperx around, because I think it's important to have not just transcription, but also timing and speaker identification (e.g. for subtitles)... It depends on pyannote, though, which has some wierd licensing (and is tougher to script the installs because of it), so I wanted to look at something that both had better transcription, and supported diarization (the speaker and timing). I decided on parakeet for the transcription with softformer (the diarization), but most of the available engines for it don't include softformer.

I coded up an OpenAI compatible server for parakeet-rs ( https://github.com/altunenes/parakeet-rs ) (which does support softformer) and I've been using it with OpenWhispr (a desktop app for transcription that handles all sorts of neat thing).

I'm doing CPU-only transcription (because I use my GPUs for other stuff and haven't gotten around to adding in the GPU-path), but it's incredibly empowering to be able to have local transcriptions at will.

This is TTS. Not STT.

For what you are doing, Senko works really well for diarization along with parakeet.

Faster and more accurate than Pyannote and whisper on my MacBook anyway.

You're right... I read the title too quickly... I'll have to look at Senko vs Softformer later...

Still I for one am grateful for any STT transcription models.

Likewise with simple whisper.cpp

A couple months back I wrote a chrome extension that does this on any webpage, with simultaneous highlighting of the sentence being read. Skips both the container launching step and the copy pasting website contents step. Might be useful to anyone trying to use kokoro ergonomically.

https://chromewebstore.google.com/detail/local-reader-ai-on-...

Love this model. I’m GPU poor and have had FOMO that I haven’t played with local models at all. About a month ago I setup Kokoro on my GTX1650 to do TTS for an article reader. A simple WebUI lets me paste a URL or a chunk of copy pasted text. Python cleans it up and sends to Kokoro for TTS and it’s then served via RSS for Apple Podcasts. Then for my morning drive I’ll catch up on articles or blog posts I’ve gathered.

At some point I’d like to play with separate voices and see if I could build something like NotebookLM for kind of like a radio morning show of news items I’ve gathered.

Check Open Notebook:

https://github.com/lfnovo/open-notebook

This repo is a good starting point for comparing TTS models https://github.com/5uck1ess/tts-bench

Kokoro is a really good model, considered it’s released 1.5 years ago. It’s punching above its weight https://5uck1ess.github.io/tts-bench/scores.html

Super cool!!

I've been using my own solution since January. I'm on Linux, and can't use Aqua, Whipsrflow etc... So i made my own.

Recently cleaned it up and made it install friendly.

If anyone is interested, you can check it out here: https://github.com/Hugo0/voiceio

It's self-improving over time, runs on your local machine, and is generally decent software. 60% of my interaction with my PC nowadays is pure voice input.

This is text to speech, yours is speech to text. (But also thanks, your STT looks interesting)

Technically Voiceio also does TTS, but it's really crappy and just meant to read stuff loud / select a lot of text and listen to it podcast style whilst I'm e.g. washing the dishes.

However you're totally right that it's focused on STT. I probably use it 95% for STT and only occasionally for TTS (which also reflects itself in the amount of polish I put into each)

and thanks!

TTS has come incredible long way, there are so many options. There is Kokoro of course, then there is Pocket TTS which is also a tiny 100M model that allows voice cloning. There is also Chatterbox Turbo, which is bit bigger but also allows for more emotional control of the voice. And then finally there is the Fish Audio S2, which is even bigger but allows even larger and essentially unbounded finegrained control of tone and emotion. And all of these can easily run on your Macbook.

Dunno why people still write about Kokoro, there are better alternatives. Pocket TTS has more natural intonation and voice cloning. Supertonic 3 is another great model in the same weight class, which is the only one that can autodetect language and make a mix of different languages sound good.

I spent a day fiddling with AI and dropping the expensive layers in kokoro, on phones, on CPU, on MNN, it runs 3x faster.

Quality is very close.

Will vary in your setup, but here is my script: https://github.com/DavidVentura/translator-rs/blob/master/sc...

I'm using Kokoro for a fun little side-project browser-based game I'm working on. It's legitimately super good for being only 85mb (for the wasm version) or 300mb (for the webgpu version).

kokoro is decent but pocket-tts is much better especially when you rip a good voice. https://github.com/kyutai-labs/pocket-tts

the onnx version of pocket-tts does perform better. https://huggingface.co/KevinAHM/pocket-tts-onnx

I've found that for CPU inference the PyTorch-based (non-quantized) version of Pocket TTS actually performs (both speed and quality-wise) better than the ONNX version, even after fiddling with all of the knobs that ONNX provides.

i found the exact opposite, the pytorch version on the cpu barely does over 2 times realtime while i can get the onnx int8 version to reach 5x.

I agree, pocket-tts is quite good. It is also very easy to make new voices from small sample files with pocket-tts.

Cool I actually got it ported to iPhone’s ANE finally yesterday! So we can get both rt natural local TTS and 4x less battery drainage and thermals

lol, I just finished a port to the iPhone ANE as well. I would love to compare notes

The biggest part was dynamic shapes that weren’t allowed for newer iOS. I’d do it with masking similar to other transformers being ported to ANE. Another thing was people would cut the graph cause some operations are not allowable on this chip - tho if you look at the op that is doing segfault, i.e. tile - you can rewrite the same functionality with other supported ops and port the weights to the new graph. Works the same way compared to PyTorch weights comparing spectrograms, around 10x realtime on 14pro :)

Anything you found interesting there? Every proposal on GitHub I found was kinda lucking in both documentation and completeness

I'm using exactly this TTS engine for my intercom door system I built. The quality of the TTS is very good.

Kokoro voices were the non-human actors directed by Opus, featured in the sound performance "God Bless the United States of Aliens". A theatrical adaptation of a liturgical texts, a surveilled mass ceremony at a protesting Christian church, being exhibited at the 61st Venice Biennale until November 2026 (Solidarity Burundi Pavilion, inside of Bosnia-Herzegovina Pavilion).

The various models of Kokoro TTS were theatrically conducted by Claude during an Anthropic hackathon. AI Voices were generated locally on a low-spec laptop to save up on API tokens and stretch the hackathon budget. Kokoro has the added benefit of being open-source.

https://daqhris.com/venice-biennale

https://daqhris.com/god-bless-usa/submission.html

I am entirely sold on Qwen3-TTS's voice cloning. It runs locally and I can run it as many times as I want.

Just tried it and it works really really well, thank you for the pointer. Came in clutch!

I built a pipeline through hermes using edge-tts to automate and listen to links that I provide to it just this morning, google notebooklm style. I replaced the TTS model with Kokoro after seeing this post, thank you. Here's the pipeline if anyone is interested. https://www.klaweht.com/2026-07-07-link-to-podcast-rss-pipel... By the way, it took hermes just around 10-15 minutes to build first iteration. I am impressed.

I get the sense you're not referring to the French fashion brand, what is Hermes?

hermes agent https://hermes-agent.nousresearch.com

similar to openclaw

thank you

you should include an example of the output of your own blog post :)

I didn't since feed address was already mentioned in the blog post. Here you go. The first one is edge-tts and the second one is kokoro. you'd be the judge :) https://ilteris.github.io/hermes-audio-feed/episodes/2026-07... https://ilteris.github.io/hermes-audio-feed/episodes/2026-07...

Naive question, but I once downloaded a particular voice file that I wanted to use with some other RVC TTS project, but ended up not being able to run it CPU only, so I only kept the voice I wanted. Thing is, the voice is in .pth format, and on Kokoro's huggingface page, their voices are all .pt.

Would I be able to use this voice I already have with Kokoro? If not, is there any way to convert it? I could always go looking if someone made this specific voice but in .pt format, but I barely mess with AI and don't know how I could search for this.

Not super familiar but it seems to be technically possible.

https://github.com/Ashish-Patnaik/kokoclone

Generate audio with the voice and your first tool and use this to clone it into kokoro.

The pth/pt extension bit isn’t the problem. Those are PyTorch extensions and they’re synonymous like jpeg/jpg.

That's great, thank you :-)

With something like this I can even try to make a more accurate voice than the one I already have, and tailor it to my liking!

I just hooked it up to my personal AI Japanese Teacher app, pretty good quality / natural sounding speech in mixed English / Japanese while running fast on CPU so I don't waste VRAM.

Was just thinking of making exactly that. I have made Kokoro the voice of my local LLM (qwen 3.6). What do you use for a Japanese tutor app?

Gemma 4 31B just because of faster MTP

For Japanese TTS, AivisSpeech-Engine[1] works really well with mixed Japanese/English text in my experience. They also provide container images on ghcr.io for both CPU and GPU inference.

[1]: https://github.com/aivis-project/AivisSpeech-Engine

Any good debian-ish distros that integrate TTS and STT in a usable shell?

Love Kokoro tts. I wrote https://github.com/Jud/kokoro-coreml to try pushing the limits a bit on speed & size. Such great quality at a given size. As others have mentioned short utterances are problematic, but solvable.

I agree that Kokoro is a good TTS model.

If you're interested in an ONNX version and a permissively licensed TTS Tokenizer, I built a pipeline for that a while back: https://huggingface.co/NeuML/kokoro-base-onnx

Saw it first on reddit, and later I created a small project to generate audio books from epub. So far I've listened to couple of books generated this way and am quite satisfied with the quality. There is just one particular word I remember that it pronounced wrongly - "Malay".

I was kind of waiting for Apple to ship something like this and using speechify in the meantime, so I was pleasantly surprised to stumble across Kokoro in a thread on HN a few days ago. The voices are better than the cloud provider IMO.

I'm using it in an open source tool I built for listening to coding agents instead of watching them work, and it's been great. Local, private, free to run, and enough voices that I can actually use a different one for each agent.

Wrote a bit more about the project here: https://jonmagic.com/posts/i-stopped-watching-my-agents-work...

You mean it's better than the speechify cloud provided voices?

Correct. In my experience so far (2 days of usage across a couple dozen voices on speechify and then Kokoro) I've preferred the ones from Kokoro.

Kokoro is great. I built an mcp for it a while back that has gotten decent traction - https://github.com/mberg/kokoro-tts-mcp if anyone wants to go that route

Both Text-to-Speech and Speech-to-Text now have local models that are good enough to get the job done. Kokoro for TTS, Parakeet for STT and Fluid-1 for text formatting (I use it with FluidVoice). I hope this is a trend that continues for other applications.

On macOS I've been using piper (https://github.com/OHF-Voice/piper1-gpl) to announce claude code notifications and it works perfectly!

It's interesting that the male voices are all so much worse than the female voices (several are quite good). There is bias in machine learning, but I wonder whether there is also systematically more training data of female speech?

I love Kokoro. I use it to read ebooks that don't have audiobooks and it works pretty good for that. I have a Python script that reads epub/html, runs the model and writes mp3s.

Kokoro supports a limited number of languages. I've had to resort to other models to support Dutch, and they just aren't as good. And usually a lot slower and bigger (too big for my 8GB VRAM). I've had to resort to the Windows built-in TTS. It doesn't sound natural at all, but at least it can pronounce most words and is very fast.

I recently created a video and didn't want to use my voice on it.

Rather than TTS I did STS (speech to speech) through Eleven Labs (you upload the video it alters the audio track to voice of choosing with some parameters you can control).

I showed it to a bunch of people and they could not believe it was non-human. The bit that got most was all the pausing, broken sentences ...essentially elements you don't put into text if your doing T2S.

An LLM over the original text could perhaps inject these elements though if prompted correctly.

I don't want to mess with port numbers and gpu cards; I just want an upgrade from the age-old:

    apt install festival

with the same installation simplicity.

I use kokoro with home assistant and its great. I find its the most natural sounding and small too.

I speak over sonos speakers when certain events happen. And use it as my voice assistant.

Did you use a Wyoming protocol wrapper to integrate with home assistant? Is it publicly available?

https://www.home-assistant.io/integrations/wyoming/

Yeah wyoming protocol I think someone even made kokoro Wyoming. I forgot exactly what I did it'd been a little while

This is awesome. People are already talking about use of local models for heavy tasks like llm inference. TTS is very much possible on device without any high end gpu. That's why Ive been working integrating it into a free reader web app with some features more than just getting audio from a text.

https://desktop.with.audio/reader

Off-topic, re: “kokoro”

Article refers to: https://huggingface.co/hexgrad/Kokoro-82M

Caught my eye for the related name to my book of Kakuro puzzles for sale at https://www.kakurokokoro.com

Kokoro comes from the Japanese word meaning something like heart or spirit, and not the literal ones.

Also from this: https://terminator.fandom.com/wiki/Kokoro

It seems all three are equally obscure. Mine at least is a heart shape :)

I wrote a CLI-based Text-to-Speech that also uses Kokoro-82M. I've been using this everyday and HN folks might find this useful too.

https://github.com/realpacific/readthis

This is Amazing and a game changer. Millions were spent during NLP era to achieve even <50% what this model offers.

Now this on a CPU is next level. When algorithms perform well on commodity hardware, the scale tips.

This gives hope that CPULLM's are not far off that'll be just fine for majority of use cases.

Just ran on Podman on a older hardware with intel CPU processor and ubuntu linux.

When given a large text, it nicely chunked them up (debug statements showed), generated the audio and played back nicely.

Well done!

Who is going to hack together a mac widget that allows us to select text anywhere, press a shortcut key and finally get a non robotic voice outputted in a reasonable amount of time?

I am aware of the Option + Esc shortcut on osx for the onboard TTS but wow is it hard to listen to in 2026.

Have you tried /usr/bin/say ? Might already have something sufficient for you - there’s quite a few voices there.

In System Settings, if you go to Accessibility and click "Read & Speak" in the "Vision" section, you can select a different voice using the "System voice" section. Click the "(i)" to preview your various options and even download more. Some, like "Allison (Enhanced)," sound leagues better than the default voice.

Great point! these are better than Samantha and they're free. But still, if I could wait a few seconds to get a much richer TTS experience I'd pay for that.

Easy to send one’s clipboard to Microsoft Azure and have their DragonHD voices read the text, say with Keyboard Maestro (or presumably Alfred, Raycast, etc.). Should work with selected text too.

You’d definitely get to pay for it, not what I consider cheap. (“$15 per 1M characters”) But IMO just about best-in-class (maybe ElevenLabs has a voice I’d like even better).

kokoro is surprisingly great at nuance but it's tough to improve that last ~2% or so. kokoro + rvc is really great too; i use that for ELEMENT47, the LLM-centric comedy podcast i do that i wish more people would listen to. (e47.net , feel free to subscribe!)

Supertonic 3 is pretty amazing too. I’ve got time to first sound down to well under a second by streaming chunks and stitching together from CPU with goonx (non c-go onnx) - happy to upload model if any go devs interested.

Interested

> Apple M2 Pro: 4.5 seconds

> AMD Ryzen 7 8745HS: 1.5 seconds

These two can probably do it much faster on their iGPUs.

Yep, and Kokoro-FastAPI (which he already uses) makes it super easy with start-gpu_mac.sh

When I hear the male voice I think: "Ok, it's the Youtube guy".

https://github.com/rhulha/StreamingKokoroJS all in browser, 100% private, nothing tracked

I tried it but it did not function.

Try it with Chrome, Firefox still has issues

Anyone know which local TTS is best, close to Eleven Labs quality?

F5-TTS

https://wlejon.github.io/kokoro-lab-web/

You can tweak the pitch as well.

this is very cool! i also made a kokoro based tts tool which runs on a jetson orin kit. it serves tts generations over durable streams, try it out here: https://streamtts.dev/ , i also wrote about it: https://s2.dev/blog/local-ai

Has anyone tested/ported Kokoro To mobile?

Another endorsement - I used Kokoro pretty extensively with an app I was developing over the last year and it's been excellent, both on- and off- GPU. Even with Elevenlabs (long time subscriber) the comparative quality of Kokoro keeps up really well until you get to their larger models with their professional voices.

I do wish there were better support for SSML, as well as deeper documentation of how to influence inflection in-line, but the default does well with standard emphasis (e.g. putting asterisks around text elements). Both asks are getting outside the zone of reasonable asks for this sort of distribution, though, and I remain incredibly grateful for the quality of what hexgrad and nazdridoy have put out in the world.

I've used Piper for local TTS before, and Kokoro looks like an interesting

the only both STT and TTS, MCP, native Mac I have found is OpenWhisperer, using it with Claude Code and Antigravity

usually TTS is not part of the package

kokoro is very nice, but I am disappointed that this wasn't an announcement of a new kokoro version.

curious to know if it comes with audio tags?

I don't want to just spew AI-hate, but is an LLM actually necessary for this? I haven't worked with, but came across loads of non-AI TTS tools. Are these now exceptionally better to justify the overhead? Genuinely asking.

It crashes as soon as you put a little paragraph of text on Apple M2 Pro.

Hard pass.

Why do these half baked projects get all the attention and thousands of clicks when it just takes a simple thing to bring the whole castle down?

Yeah, we need to keep up with how quickly AI types back to us, typing on the keyboards is no longer quick enough, gotta dictate everything now.

This is the opposite way

Great way to enter your passwords

Reply on news.ycombinator.com