4 ms·

You can get pretty close with open source software: https://claudio.uk/posts/audiblez-v4.html https://claudio.uk/posts/audiblez-v4.html

rapind 2y ago
Oh wow. Thanks for posting! Samples sound great (on par with eleven by my untrained ear). Will definitely use this.
neom 2y ago
How does it hold up on long stuff? I use Elevenlabs Studio daily and once things start to get into the chapters long, the voice can really start to go off the rails. It'd say they've solved a lot of this over the past 2/3 months, but it does still happen on long stuff.
- masteruvpuppetz 2y ago
  >> the voice can really start to go off the rails. Do you mean the AI gets tired?
  - zaptrem 2y ago
    In autoregressive models error accumulates over time. He likely means the voice starts to make odd sounds/gets lower quality. It would be really interesting if OP could share a clip of this phenomenon!
    - neom 2y ago
      Various different things can happen, it would take me quite some time to dig up examples but at least with elevenlabs you don't get the clicks and pops you get like on notebook LM for example. 11labs instability comes in the forms of intonation, pitch, accent, garbled words or even once language. I've only seen it happen in the 3k+ words gen's I've done, usually actually around the 75% point of the narration of whatever I've converted, and on average lasting a couple of seconds top.
      - wrsh07 2y ago
        Yeah - I've experienced this with eleven reader (I don't think you can gen text this long anymore using the reader app, lol) but switching voices fixed it for me I can go back and try to repro and get a recording....
- csantini 2y ago
  It holds up well, because Audiblez uses sentence splitting (via Spacy models) before audio synthesis
ultrasounder 2y ago
Bravo!
simongray 2y ago
Oh no, it doesn't run on Apple Silicon. That's too bad.
- _joel 2y ago
  > On my M2 MacBook Pro, on CPU, it takes about 1 hour, at a rate of about 60 characters per second. Umm, it does.
  - simongray 2y ago
    My bad. I misread the official website: > We don't currently support Apple Silicon, as there is not yet a Kokoro implementation in MLX. As soon as it will be available, we will support it. I thought that meant that it didn't support Apple Silicon in general, but they were just talking about GPU support.
    - fl0id 2y ago
      though they wouldn't need to use MLX, could also use pytorch etc
      - stoobs 2y ago
        I think there's an issue somewhere in Kokoro though which means it doesn't actually take advantage of MPS, I did get a modified version up and running, but it was no faster than CPU, even though it passed all the internal tests using mps. I might try using F5-TTS-MLX instead actually (https://github.com/lucasnewman/f5-tts-mlx https://github.com/lucasnewman/f5-tts-mlx) and see how that does.
- csantini 2y ago
  It works on Apple Silicon, but it doesn't use the GPU. Because Kokoro has not been implemented yet in MLX
  - simongray 2y ago
    Ah my bad! I just read the "We don't currently support Apple Silicon" on the official website, but I didn't realise that only pertains to GPU support.
tonyhart7 2y ago
good, now how I can use this on mobile??
- csantini 2y ago
  Generate the audiobook on a laptop and then listen to it on mobile
  - tonyhart7 2y ago
    this is the easy way, but I want the hard way
laurentlb 2y ago
Interesting! This uses the Kokoro-82M model, which has a pretty good quality, but the set of languages is still quite limited.
anonymous344 2y ago
does this run on linux machine also?
- nkmnz 2y ago
  third line on the page right below the first image says: > Audiblez 4.2 running on MacOSX via wxWidgets. Linux and Windows are supported too

rapind 2y ago

Oh wow. Thanks for posting! Samples sound great (on par with eleven by my untrained ear). Will definitely use this.

neom 2y ago

How does it hold up on long stuff? I use Elevenlabs Studio daily and once things start to get into the chapters long, the voice can really start to go off the rails. It'd say they've solved a lot of this over the past 2/3 months, but it does still happen on long stuff.

masteruvpuppetz 2y ago
>> the voice can really start to go off the rails. Do you mean the AI gets tired?
- zaptrem 2y ago
  In autoregressive models error accumulates over time. He likely means the voice starts to make odd sounds/gets lower quality. It would be really interesting if OP could share a clip of this phenomenon!
  - neom 2y ago
    Various different things can happen, it would take me quite some time to dig up examples but at least with elevenlabs you don't get the clicks and pops you get like on notebook LM for example. 11labs instability comes in the forms of intonation, pitch, accent, garbled words or even once language. I've only seen it happen in the 3k+ words gen's I've done, usually actually around the 75% point of the narration of whatever I've converted, and on average lasting a couple of seconds top.
    - wrsh07 2y ago
      Yeah - I've experienced this with eleven reader (I don't think you can gen text this long anymore using the reader app, lol) but switching voices fixed it for me I can go back and try to repro and get a recording....
csantini 2y ago
It holds up well, because Audiblez uses sentence splitting (via Spacy models) before audio synthesis

ultrasounder 2y ago

Bravo!

simongray 2y ago

Oh no, it doesn't run on Apple Silicon. That's too bad.

_joel 2y ago
> On my M2 MacBook Pro, on CPU, it takes about 1 hour, at a rate of about 60 characters per second. Umm, it does.
- simongray 2y ago
  My bad. I misread the official website: > We don't currently support Apple Silicon, as there is not yet a Kokoro implementation in MLX. As soon as it will be available, we will support it. I thought that meant that it didn't support Apple Silicon in general, but they were just talking about GPU support.
  - fl0id 2y ago
    though they wouldn't need to use MLX, could also use pytorch etc
    - stoobs 2y ago
      I think there's an issue somewhere in Kokoro though which means it doesn't actually take advantage of MPS, I did get a modified version up and running, but it was no faster than CPU, even though it passed all the internal tests using mps. I might try using F5-TTS-MLX instead actually (https://github.com/lucasnewman/f5-tts-mlx https://github.com/lucasnewman/f5-tts-mlx) and see how that does.
csantini 2y ago
It works on Apple Silicon, but it doesn't use the GPU. Because Kokoro has not been implemented yet in MLX
- simongray 2y ago
  Ah my bad! I just read the "We don't currently support Apple Silicon" on the official website, but I didn't realise that only pertains to GPU support.

tonyhart7 2y ago

good, now how I can use this on mobile??

csantini 2y ago
Generate the audiobook on a laptop and then listen to it on mobile
- tonyhart7 2y ago
  this is the easy way, but I want the hard way

laurentlb 2y ago

Interesting! This uses the Kokoro-82M model, which has a pretty good quality, but the set of languages is still quite limited.

anonymous344 2y ago

does this run on linux machine also?

nkmnz 2y ago
third line on the page right below the first image says: > Audiblez 4.2 running on MacOSX via wxWidgets. Linux and Windows are supported too