Tamas G

Aha, what happens when you try to do weird experiments on a Friday night and land on, "let's see what Speechbox driven through a tube model might even sound like at the most basic level!"
Like, super crude. a squared sine for the open phase and an exponential decay for the return. Basically the minimum
viable "not a square wave" pulse shape. What do you get?
Why of course, electric buzzer and misquito buzzer! I give you, 6 vowels. Let's play a new gameshow. Can you tell the vowels within the buzz? Ha. @fastfinge@fed.interfree.ca would be amused at this one. And no, we're not going tube model. 16 times the processing power at 44100 because the tube physics requires it is just. brutal.

there you go you fluffy little testflight peoples, you get a new build with the ever-evolving editor tab, year-digit-pair splitting, your breakfasts are fixed from now on, and dates correctly have cardinal numbering support. Honestly a big one, now I am super sleepy from this and all the at-work things doing documentation and such.

Lunch time! You know what that means! New TestFlight build for testers with some easy wins! Let's see if I can crank up some commits real quick before the work swamp swallows me again.

lol now someone found an edge case for the shimmer issue on diphthongs, and I can't un-hear it. Type the word how, just on its own. How. And slow it between rates 55 to 65. This creates some of the same shimmery effect I tried to solve. Oh no. Back to the drawing board. But then, in "how are you?" it's not really there even at that rate. Oof. Now my ears can't unhear it though.

for now. to fix your breakfast after updating to 3.0 Beta 4 in NVDA, type this into your run dialog:
%appdata%\nvda\addons\tgspeechbox\synthdrivers\tgspeechbox\packs\dict\en-compounds.tsv
And look for the word breakfast, remove that one and the under it which is breakfasts. Poof. Once those two lines are gone, it won't get split anymore. Solidly a good catch by @danestange@caneandable.social!

TGSpeechBox v3.0-beta4 is out — 43 commits, 12 bug fixes, 8 new features, 11 language pack improvements.
Biggest addition: 3,686 compound words now split before phonemization and merge back seamlessly. "Popcorn" gets correct vowel quality on both halves. Combined with new word boundary amplitude dips, "lighthouse keeper" vs "light housekeeper" now sound properly different.
Diphthong shimmer — the bug that survived three rounds of fixes — is finally dead. Staircase micro-frames with discrete formant steps replace the continuously-changing targets that IIR resonators couldn't track cleanly.
Numbers now expand via YAML rules, fixing stress alignment across every platform. Impulse pitch gets real terminal gestures — declaratives fall, questions rise.
Spanish gets a community-driven tuning round from @rmcpantoja@mastodon.social. Polish phoneme work started thanks to @spacepup@mastodon.stickbear.me and @pitermach@dragonscave.space. GenAm gets proper MOUTH onset from Hillenbrand data and -og words like "dog" shift LOT→THOUGHT where they belong.
GOAT gets its diphthong glide back in en-us too. Turns out monophthongs were not the way.
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/TGSpeechBox-v300b4.nvda-addon
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/TGSBPhonemeEditor-v300b4.zip
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/TGSpeechSapiSetup-v300b4.exe
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/TGSpeechBox-v300b4.apk
https://testflight.apple.com/join/jvvGY6Fz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/tgspeechbox-linux-x86_64-v-300b4.tar.gz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b4/tgspeechbox-linux-aarch64-v-300b4.tar.gz

What makes me a little sad about AI is just how coders will know less about what code even means over time, and we'll forget the exact commands to build ourselves, because an AI can run those commands. What's a constructor again and how does it relate to creating a new object and member in a class? When is the value passed in on the handle once during initialization versus as a callback dispatched between parts of your app? Coders aren't going to know these terms anymore possibly. They'll just roll their eyes and say, "well, the AI can choose for me." and that's going to create so much buggy, memory-schlepping software, our CPUs won't know what to do with it. Coders who are dropped into a terminal will not know how to even run a simple build command for a project without their AI code monkeys helping them. Blique future indeed.

yes I dogfood my own Speechbox. Every day, all the time. Sometimes at rate 60, sometimes rate 70, sometimes rate 80. Yes I'm aware there's a shimmer effect on the word "how" at certain rates when on its own, and its a limitation of just how the vowel jump happens in the cascade resonators. I spent 2 hours and 14 different ways to debug it at night, didthering, F1-F2 shifting at the onset, adding a buffered aspiration before the diphthong glide, all of it. I know my own architecture and am actively refining it, don't you all worry, the fixes are coming. Just give it time. British will improve on its glides, word-level dictionary for text replacements, all of that is in the works. But it can't come over night.

how how. External Group ∙ 76 Testers ∙ 10 Builds what the number keeps climbing y'all. I'm scared. haha.

lol. Seeing things like this in articles: "Why did we need to attach “AI-Native” to 6G? You know why. Here’s to hoping AI is dead by 2029."
As well as people saying how AI will die once the bubble bursts, how it will go away, never to be seen again or to only be used in niche cases. Boy will you be surprised when the AI bubble pops but AI investment and pace only slows down, not sputters out. Did the internet collapse after the dot com bubble burst? Nope. It's still here. Why people think AI is any different is beyond me. It's staying around, whether you like it or not. Enough of the world has seen more benefits, particularly the financially-motivated people at the top, that nobody's stuffing it back in the bottle even with a market collapse. Sure the companies will be worth less, hold less power, that will be good overall, but AI, as a concept, as a thing, will go nowhere.

So let me get this one straight in my head. When VoiceOver swipes fast, it fires cancelSpeechRequest then immediately synthesizeSpeechRequest for the next element. Our cancel clears the buffer under the mutex, but if the render block is mid-flight, it might still be reading stale data. I snooped around to see what was done to solve this, especially as Espeak-NG-Mobile licensed their code under MIT, and
Maybe I can instead replace the semaphore with os_unfair_lock, and in the render block use os_unfair_lock_trylock() instead of a blocking wait. If the lock is currently held by cancel or synthesize, trylock returns false immediately, no waiting. The render block just fills the buffer with silence for that one callback (~2-5ms) and returns. That tiny silence gap is completely inaudible. But a stalled audio thread that misses its deadline? Better to not have that.

oh holy crap I'm loving this already.
Anthropic just released a mobile version of Claude Code called Remote Control
https://venturebeat.com/orchestration/anthropic-just-released-a-mobile-version-of-claude-code-called-remote

TGSpeechBox v2.97 is out!
The big one: diphthong collapse. Diphthongs used to render as two separate vowel beats stitched by a crossfade. Now they merge into a single token with cosine-smoothed micro-frame waypoints — one continuous formant sweep. Pitch-adaptive frame intervals keep the glide clean across voice profiles.
Two new pitch models. Impulse style: warm, naturally decaying stress peaks with IIR smoothing — first stress gets the biggest bump, each one after it fades. Klatt style: the classic 1987 hat pattern — rise on first stress, plateau, terminal fall for statements or rise for questions. Both join espeak_style, legacy, and fujisaki_style.
Spanish got a major upgrade thanks to Mateo Cedillo. Both Castilian and Mexican packs now have proper lenition, cluster separation, tuned five-vowel triangle, and dialect-specific sibilants. Mexican seseo is now bulletproof regardless of eSpeak version.
Radiation model overhaul: lip radiation moved from pre-cascade to post-cascade. The old model boosted upper formants giving that "buzzy Klatt" character. The new additive model preserves the cascade's natural formant balance — warmer vowels with proper chest voice weight. Thanks to Simon (@simon@procrastodon.net) for flagging the spectral gaps that sparked this.
https://github.com/tgeczy/TGSpeechBox/releases/download/v-297/tgSpeechBox-v297.nvda-addon
https://github.com/tgeczy/TGSpeechBox/releases/download/v-297/TGSBPhonemeEditor-v297.zip
https://github.com/tgeczy/TGSpeechBox/releases/download/v-297/TGSpeechSapiSetup-v297.exe
https://github.com/tgeczy/TGSpeechBox/releases/download/v-297/tgspeechbox-linux-x86_64-v-297.tar.gz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-297/tgspeechbox-linux-aarch64-v-297.tar.gz

When legacyPitchMode: "klatt_style" is enabled, these settings control the intonation:
Hat rise and stress peaks:
• klattHatRiseHz (number, default 30.0): Pitch step-up (Hz) when entering the hat plateau on the first primary-stressed vowel. Scaled by inflection.
• klattStress1Hz (number, default 28.0): Additive pitch peak on the 1st stressed vowel on the plateau.
• klattStress2Hz (number, default 16.0): Peak on the 2nd stressed vowel.
• klattStress3Hz (number, default 13.0): Peak on the 3rd stressed vowel.
• klattStress4Hz (number, default 11.0): Peak on all subsequent stressed vowels.
Declination and terminal gestures:
• klattDeclinationHzPerSec (number, default 10.0): Linear baseline declination rate.
• klattFinalFallBelowBaseHz (number, default 21.0): How far below baseline the pitch drops on the hat fall for statements/exclamations.
• klattQuestionRiseHz (number, default 35.0): Pitch rise on the last stressed vowel for questions (instead of a fall).
• klattContinuationRiseHz (number, default 15.0): Pitch rise for commas/continuations.
• klattGlottalLowerHz (number, default 15.0): Additional pitch drop on the very last vowel for statements/exclamations. Simulates glottal lowering at utterance end.
Smoothing:
• klattSmoothAlpha (number, default 0.4): Single-pole IIR smoothing coefficient (0..1). Lower = smoother transitions between hat

well, I was up until 1:30 for a good reason, I swear! Now added a new Diphthong collapse pass. Over 400 lines of code to do proper micro-frame emitions for glides, splitting them into up to 10 segments with cosine-smoothed formant waypoints. I had initially started with 3 frames max no matter what, but now realized it has to be between 3 and 10 depending on the glide's duration, with shorter glides needing more micro-frames to not have a jitter effect. Solved it well to where we can now rely on Espeak's Diphthong output along with auto-tieing them for vowel to vowel pairs (but not semivowel to vowel, that was a good bug to find!)
the merged token's duration is divided by microFrameIntervalMs (default 12), clamped to 3–10
frames. So an 80ms diphthong gets ~7 micro-frames of ~11ms each. A 150ms diphthong gets 10 frames of 15ms each.
So yes, up again, late, for a good reason, at least to me. What a journey into DSP mechanisms and micro frames.

It's so weird with Claude code, the more demanding and frustrating I get with it to correctly locate the bug, fix the right code path, don't miss lines in such and such file, the kinder it gets to me, LOL.
You would think it would be like, "don't be so demanding!" but it's like, "oh, yeah, I get it, you might be demanding to me about this but I won't take it is an attitude from you." LOL. I find that amusing in some way.
As a machine, maybe this is where it has an advantage, it will never read tone from human writing the same way a human does, so of course it's not going to know whether what I write is really demanding or just being precise, to it the same generalities apply. If I were working with another paired engineer I would definitely have to dial that back and be a lot more colaborative as they work through each step too, but with Claude that means interrupting the thinking chain mid-process which isn't always good. So yeah, completely different approaches on all levels to working with a paired engineer, LOL.

wow, merged the first real quality-of-life improvements into ES.yaml and the phonemes. Spanish is about to sound a lot more Spanish :) This person (Mateo) did a great job adding new language phonemes and fighting some Espeak rules too to sound better.

And that was all knocked out in like 40 minutes of work, 20 of which was manual writing, 10 for release notes I typed up and then draft converted, get the Claude skill briefing done for packaging releases, which I won't have to again. Not bad. It's not a Claude skill, but I'm sure people have turned this into one by now.
If it has step by step markdown on what to do it does really good at splitting each stage into tasks and then running down them to remove the recursive grind. I think 10 years ago I would have gotten the same though by writing a bunch of batch files and scripts. All of this work can be automated without AI, just not in the form of natural language. It's not as impressive when I know that yes, if you took me back to 2016 as who I am today, I would have found another way 100%.

TGSpeechBox v2.96 is out!

The headline fix: a cascade corruption bug in the normalization engine was mangling language packs that remap phonemes with suffixed variants (a to a_es). Each rule's output fed unprotected into next rules, garbling the stream. Found by Mateo during Spanish pack development. Fixed with length-aware protection that blocks corruption while preserving intentional chaining.
Quality: the frame emission pipeline now generates micro-event sub-segments -- voice bars, bursts, aspiration, fricative attack/decay -- each with independent timing. Stop onsets are cleaner, affricates transition properly.
Duration got a serious upgrade. Voiceless codas lengthen to compensate for pre-voiceless vowel shortening. Phrase-final nasals sustain naturally. Stops scale by place -- velars longer than alveolars, matching how articulators move.
Consonant identity is sharper: five nasals now have distinct spectral signatures via differentiated anti-resonance frequencies. Velars got context-dependent coarticulation -- different F2 loci before front vs back vowels. This was the missing piece that made "go" and "doe" sound identical.
Next release: Spanish gets separation from sounding less Castilian in ES-MX, and other tuning will get merged, since the PR is close, further tuning on languages like english "dot" and "fox" which are too short thanks to Espeak.
https://github.com/tgeczy/TGSpeechBox/releases/download/v-296/tgSpeechBox-v296.nvda-addon
https://github.com/tgeczy/TGSpeechBox/releases/download/v-296/TGSBPhonemeEditor-v296.zip
https://github.com/tgeczy/TGSpeechBox/releases/download/v-296/TGSpeechSapiSetup-v296.exe
https://github.com/tgeczy/TGSpeechBox/releases/download/v-296/tgspeechbox-linux-x86_64-v-296.tar.gz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-296/tgspeechbox-linux-aarch64-v-296.tar.gz

lool pretty awesome, in 15 minutes time I wrote up a markdown briefing to Caude on how to publish my releases and package consistent fresh files in each asset, rebuild them, put them up. Now combined with the CI actions this gets very powerful, from start to finish a release can be put in draft and I just verify it looks right, assets are fresh and boom. Way less steps to any release. It makes releasing on a lunch break more possible, which is what I'll do

TGSpeechBox v2.95 is out!
The old "rate reduction" pass just shortened everything equally at high speed — schwas, stops, fricatives all got crushed the same way. At 3x rate, segments became acoustically dead. v2.95 replaces it with rate compensation: per-class duration floors that protect intelligibility at speed. Vowels, fricatives, nasals, stops — each has its own minimum. A cluster proportion guard prevents timing bulges when one consonant hits its floor. At normal speed it does nothing.
Trajectory limiting got retuned (cf2 18->26 Hz/ms fixes the "Shaggy voice" diphthong smear) and now supports full nested YAML with bitmask field selection.
The phoneme editor had 9 bugs from its first third-party use. All 9 are fixed: surgical save preserves comments and key ordering, flow maps no longer corrupt voice profiles, and a new scope awareness dialog scans all 27 language packs before letting you modify a shared phoneme — offering to clone it for your language instead. Flag change and formant ordering validation round it out.
Polish gets microprosody, trajectory limiting, and rate compensation tuned for its dense clusters. The /i-bar/ vowel got a major formant correction backed by published research.
TGSpeechBox now crosses 270 tunable per-pack settings, all localizable in the NVDA driver.
https://github.com/tgeczy/TGSpeechBox/releases/download/v-295/tgSpeechBox-2026-v295.nvda-addon
https://github.com/tgeczy/TGSpeechBox/releases/download/v-295/TGSBPhonemeEditor-v295.zip
https://github.com/tgeczy/TGSpeechBox/releases/download/v-295/TGSpeechSapiSetup-v295.exe
https://github.com/tgeczy/TGSpeechBox/releases/download/v-295/tgspeechbox-linux-x86_64-v-295.tar.gz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-295/tgspeechbox-linux-aarch64-v-295.tar.gz