@ZDi____ profile picture
ZD1908 @ZDi____

🇦🇷 25M ; ML Text-to-Speech/Audio, C /Qt ; DMs open

Joined June 2024
  • Tweets 3,872
  • Following 442
  • Followers 547
642 Photos and videos
Pinned Tweet
ZD1908
@ZDi____
Jun 13
Finetuned MOSS-TTS (the 1.7B version) on some speakers, then trained a new decoder from scratch, based on iSTFTNet3 that takes in its tokenizer features and outputs 48KHz audio. I am a genius. I'll open source this later.
2:00
19
7
200
13,109
Good afternoon.

0:53
41
I was using Le Chaton Fat to start another TTS training run. When it saw my dataset, it uninstalled itself. AGI.
2
25
812
You can take a boy out of 4chan, but you can never take the 4chan out of a boy.
1
12
174
ZD1908
@ZDi____
Jun 14
If Codex made a car I made this out of frustration because Codex loves to put incessant labels on everything when writing frontend.
1
1
7
494
Gm frens 👹🥖 “Victory belongs to the most persevering.” Napoleon Bonaparte
0:16
3
38
660
ZD1908
@ZDi____
Jun 14
It's strange how Elon Musk is the world's first trillionaire... and he doesn't have much political influence? If I were a trillionaire, San Francisco would be littered with statues of me and the entire tech elite would answer to me. X would run the US govt like Samsung does SK's
3
7
219
ZD1908
@ZDi____
Jun 14
The government and regulatory system would be so obedient to me that they'd let SpaceX throw a thousand endangered species into the sun for more rocket experiments if we wanted to.
1
126
ZD1908
@ZDi____
Jun 14
Where were you when Clavicular got brutally framemogged by the ASU frat leader?
2
242
ZD1908
@ZDi____
Jun 13
My mom goes around telling her friends how much I make (very good relative to this country's mean wages), and some of them ask if I need a girlfriend. I do not want a girlfriend. I am a firm trvcel chud and will keep my purity.
1
5
223
ZD1908
@ZDi____
Jun 13
Finetuned MOSS-TTS (the 1.7B version) on some speakers, then trained a new decoder from scratch, based on iSTFTNet3 that takes in its tokenizer features and outputs 48KHz audio. I am a genius. I'll open source this later.
2:00
19
7
200
13,109
ZD1908
@ZDi____
Jun 13
Neural vocoder has ~31M parameters. All on 1xMI300X thanks to @HotAisle and @AIatAMD
1
18
937
Gm frens👹🥖 “One against many is not a disadvantage , it is an opportunity to prove that one man can be worth an army.” Napoleon Bonaparte
0:15
2
10
77
1,899
ZD1908
@ZDi____
Jun 13
Indians are not only good at IT, but they can manufacture very good bikes. My Bajaj Dominar 400 has always run smoothly.
ZD1908
@ZDi____
Jun 12
Wake up, go out on my bike, eat milanesa and fries at a bodegón. Life is good
2
7
586
ZD1908
@ZDi____
Jun 13
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA I HAD NEAR SOTA AUDIO AND TTS MODEL IN AN INSTANCE WHICH THEN DIED
2
114
ZD1908
@ZDi____
Jun 13
Why are Russian artists so good? Is it something they put in the vodka? I almost find myself cursing Putin for making Russia a pariah on the international stage.
1
1
177
ZD1908
@ZDi____
Jun 12
Wake up, go out on my bike, eat milanesa and fries at a bodegón. Life is good
2
3
1,309
ZD1908
@ZDi____
Jun 12
I've been training my iSTFTNet3-48KHz as a MOSS Audio tokenizer decoder and the audio quality I got on my small dataset is almost perfect. Results tomorrow.
20
1,270
ZD1908
@ZDi____
Jun 12
Testing something Uhhh. SpaceX
201
ZD1908
@ZDi____
Jun 11
I am not spending enough money on art commissions.
1
246
ZD1908
@ZDi____
Jun 11
It works! I made it a decoder-only TTS model. Text then mel with the head. This one is just 52M param, trained from scratch on LJSpeech (20 hours). The audio quality is shitty because I'm inverting the melspectrogram with Griffin-Lim.
0:20
ZD1908
@ZDi____
Jun 10
I'm running an experiment. With AR transformers for speech, do we need a tokenizer, or can we get away with predicting mel spectrogram directly? This unconditional transformer predicts a latent which then goes into a 1D causal conv that predicts the next 4 mel frames.
12
13
305
54,104
Load more