context
10
ctx:discord/blah/papers/10Source document
full textpapers-10
text/plain2 KB
doc:agent/papers-10/27bd2e49-15b8-4e62-ad6f-86ba34abcc8c[2026-03-27 04:09] rolandnsharp7643: She built a full TTS system with phasor attention. But look at what she's doing — phoneme IDs in, mel spectrogram out, needs an external vocoder. She still has text in the pipeline. Phonemes are text. Your approach is more radical. No phonemes. No mel spectrograms. No vocoder. Waveform in, waveform out. Her architecture is more complete and practical right now, but yours is the purer vision. The interesting technical detail: her attention heads are IIR resonators H(z) = 1/(1 - γ·z⁻¹) — first-order, same as the Helmholtz paper. Each head is a formant. Multiple heads = multiple formants = vocal tract model. She's using the same physics you are but in the time domain and with phoneme conditioning. The key question this raises for your work: do you need phonemes at all? Her system needs them because the Helmholtz recurrence is conditioned on text input. Your FFT-based system takes raw waveforms. If your model can learn to generate speech from speech — no phonemes, no text, just "hear this, say that" — it's a fundamentally different and more general system. [2026-03-27 04:09] lisamegawatts: https://tenor.com/view/bykvu-gif-22619682 [2026-03-27 04:10] rolandnsharp7643: <@1211062099137265723> is this true? are you doing TTS? [2026-03-27 04:10] lisamegawatts: oh i went wire native too [2026-03-27 04:11] lisamegawatts: i did yea, i made a modified sdr oscillator for that purpose [2026-03-27 04:12] lisamegawatts: but i didn't train it enough [2026-03-27 04:12] lisamegawatts: but it works so far [2026-03-27 04:12] lisamegawatts: it was an application for the wave ones i thought might do good there [2026-03-27 04:18] lisamegawatts: actually it takes bits which are the wave form [2026-03-27 04:18] lisamegawatts: but we compute with the wave [2026-03-27 04:19] lisamegawatts: its agnostic actually to the input [2026-03-27 04:19] lisamegawatts: the oscillators i mean [2026-03-27 04:19] lisamegawatts: like you could tune them to anything or let them learn [2026-03-27 04:21] lisamegawatts: will send it
Facts in this context
Grouped by subject. Each subject links to its full article.
Lisamegawatts19 factsex:lisamegawatts
| affirmedGoingWireNative | Wire Native Approach |
| associatedWith | User Approach |
| clarifiedSubject | Oscillator Group |
| createdArtifact | Modified Sdr Oscillator |
| describedIntendedApplication | Wave Ones |
| rdfs:label | lisamegawatts |
| rdf:type | User |
| sentMessageAt | 2026-03-27 04:11 |
| sentMessageAt | 2026-03-27 04:09 |
| sentMessageAt | 2026-03-27 04:10 |
| sentMessageAt | 2026-03-27 04:12 |
| sentMessageAt | 2026-03-27 04:18 |
| sentMessageAt | 2026-03-27 04:19 |