GameCult/AquaSynth
Watch
9
0
Fork
You've already forked AquaSynth
 0
Code Issues Pull requests 4 Projects Releases Packages Wiki Activity Actions

Clarify PanPhon feature basis

Browse source
This commit is contained in:
Meta Vens 2026-05-21 10:21:47 +01:00
commit 641989db12
2 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

2
docs/ipa-vocal-tract-roadmap.md
View file

@ -247,7 +247,7 @@ AquaSynth should treat it as the training input contract:
The first Weksa v0.1 training artifact now exists at
`E:\Projects\weksa\examples\speech-training\tiny-panphon-v0.1\batch.json`.
It covers the crawl-stage IPA set `a`, `pa`, `ta`, `ka`, `sa`, and `ma` with
fixed vector widths, zeroed semantic text embeddings, inline PanPhon-style
fixed vector widths, zeroed semantic text embeddings, inline PanPhon 22-feature
sequence evidence, deterministic prosody hints, and an Epiphany-compatible
neutral character-state vector. Toy vectors are allowed for plumbing tests only.
Do not feed IPA strings into the semantic text embedding channel and pretend

2
state/evidence.jsonl
View file

@ -124,4 +124,4 @@
{"date":"2026-05-20","topic":"speech-training-time-witness","lesson":"Speech training receipts need time and confidence as evidence, not only loss. Each curriculum step should record when the stage decided, observed latency, intended latency budget, confidence, and artifact/model identifiers; a good-sounding result from an unbounded clockless path is not trustworthy.","evidence":"Added SpeechTimingReceipt to SpeechRenderRequest, SpeechRenderResult, and SpeechTrainingCheckpoint; seeded timing/confidence receipts in the local distributed-training proof; updated audio doctrine, IPA roadmap, utterance parity research notes, spine, and memory."}
{"date":"2026-05-20","topic":"weksa-utterance-embedding-contract","lesson":"Do not start serious AquaSynth speech gradient descent from toy Weksa vectors, and do not treat semantic text embeddings as IPA embeddings. The coordinated Weksa handoff contract is `weksa.utterance_embedding_handoff.v0.1`: 1024 `bge-m3:latest` semantic speech-text floats, variable-length PanPhon sequence evidence, 256 AquaSynth-owned phonetic-realization floats produced by `aquasynth.panphon_sequence_encoder.v0.1`, 32 prosody/emphasis floats, 64 projected character-state floats, and a 64-float AquaSynth-owned utterance embedding output.","evidence":"Updated E:/Projects/weksa docs/schema/state to split speech text embedding from PanPhon sequence evidence and marked the existing Nibu toy fixture as non-training; mirrored the training block in AquaSynth IPA roadmap, utterance parity research notes, spine, and memory."}
{"date":"2026-05-20","topic":"panphon-sequence-encoder-organ","lesson":"The 256D phonetic realization surface is now an AquaSynth-owned learned organ, not fixed phone slots. Weksa supplies variable-length PanPhon feature sequences; AquaSynth encodes them through a trainable phonetic sequence encoder before the utterance embedding encoder consumes the 256D result.","evidence":"Added PhoneticSequenceEncoding.cs with PanPhon feature frames, deterministic sequence sketching, PhoneticSequenceNeuralEncoder, training records, and tests proving the encoder learns compact PanPhon realization embeddings."}
{"date":"2026-05-20","topic":"weksa-v01-training-artifact","lesson":"The Weksa blocker has moved from schema definition to artifact consumption. Weksa now emits a machine-readable `weksa.utterance_embedding_handoff.v0.1` tiny PanPhon seed batch for `a`, `pa`, `ta`, `ka`, `sa`, and `ma`; AquaSynth's next serious training pass should load that artifact rather than local toy vectors.","evidence":"Added Weksa emitter `E:/Projects/weksa/tools/emit_tiny_panphon_handoffs.py` and generated `E:/Projects/weksa/examples/speech-training/tiny-panphon-v0.1/batch.json` with fixed 1024/32/64 input widths, inline PanPhon-style frames, and pending AquaSynth-owned 256/64 learned outputs."}
{"date":"2026-05-20","topic":"weksa-v01-training-artifact","lesson":"The Weksa blocker has moved from schema definition to artifact consumption. Weksa now emits a machine-readable `weksa.utterance_embedding_handoff.v0.1` tiny PanPhon seed batch for `a`, `pa`, `ta`, `ka`, `sa`, and `ma`; AquaSynth's next serious training pass should load that artifact rather than local toy vectors.","evidence":"Added Weksa emitter `E:/Projects/weksa/tools/emit_tiny_panphon_handoffs.py` and generated `E:/Projects/weksa/examples/speech-training/tiny-panphon-v0.1/batch.json` with fixed 1024/32/64 input widths, inline PanPhon 22-feature frames, and pending AquaSynth-owned 256/64 learned outputs."}