I ended up fixing a bunch of long-standing low-level bugs that were responsible for pops, clicks, and general weirdness. The noise generator had DC bias, filters were carrying energy across utterances, frame data wasn’t fully initialized, and there was even a one-sample “zombie frame” when going from silence to speech. All of that is gone now. Even with fast speech, letter echo, or lots of interruptions, it stays clean.
The other big thing is trills. Rolled R’s are now handled in the engine instead of through pack hacks like doubling letters. There are two pack-level settings now: one that controls how long the trill lasts, and one that smooths it. The actual flutter speed is fixed to something that matches how real trills behave, so pack authors don’t have to fight the engine anymore. Short values behave like taps, longer values give you a proper roll, and it’s consistent across languages.
I tried pushing the engine toward a more “rounded” glottal pulse as well, but that changed the character too much for now, so I backed it out. The goal here was stability and correctness first, not a surprise timbre shift. Now that the engine isn’t fighting itself anymore, future tuning should be much easier and safer.
Language updates: Polish, US English says words like "start", "neat" and "need" more correctly.
Download: eurpod.com/synths/nvSpeechPlay…
The onslaught includes LLMs finding bogus vulnerabilities and code that won't compile.
arstechnica.com/security/2026/…
I am strongly considering shutting down Pomf because the US Department of Justice continues to manufacture outright lies against people who are innocent until proven guilty.
Running the service (despite my best efforts to mitigate risks well beyond what the law would require me to do up to and including full fledged cybersecurity research) exposes me to some level of legal or criminal threat to my livelihood. I knew this going in five and a half years ago, and the calculation at the time was acceptable because even if the feds came knocking at my door, I was confident the evidence would be in my favor and that a reasonable and functional court system would make the right decision. I was also confident that what happened to Les De Ridder almost eight years ago (archive.is/PJTzS) wouldn't happen here in the USA, because Europe was a communist shithole and we had rights over here.
Well, as the US continues to backslide into a fascist regime with a completely captured judicial branch that is utterly subservient to the executive branch, my evaluation of that risk level compared to my maximum tolerance of risk continues to inch closer and closer to parity, and when that risk exceeds it, I am out. If I am not convinced that I can adequately defend myself against potentially spurious claims and threats due to a corrupt and unequal justice system, then my next defense mechanism is to remove any and all ammunition from those who would try to harm me, and the largest weapons cache someone can bring to bear against me at this time is undoubtedly Pomf. I have never considered the government to be part of my overall threat model, but now I do, and I do not have the energy or resources to fight an entire government at this time despite it being morally the right thing to do.
If this does happen, there will be a reasonable and well defined sunset period, with a final archiving of all Pomf content to cold storage in hopes that in the distant future the risk comes down to a level in which I am comfortable bringing it all back online. I would never wipe Pomf - only make it unavailable at worst.
One last thing - if you think I am some big baby or think I have nothing to fear, consider this simple statement:
If they can do it to them, they can do it to you.
Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation
I haven't been paying much attention to the state-of-the-art in speech generation models other than noting that they've got really good, so I can't speak for how notable this new …Simon Willison’s Weblog
Arfy
in reply to Rosalyn Anne • • •