babihoba/supertonic3-optimization

Supertonic 3: F1200 single-thread, eight-step student

Selected model: models/round5-original50-student/. The user selected F1200 after comparing it with E400 and hearing no large difference on the supplied Korean development listening samples. This is limited listening feedback, not a MOS score or a perceptual similarity percentage.

F1200 uses unchanged FP32 duration/text/vocoder components and reduced-range dynamic INT8 for the vector estimator. Inference uses one CPU thread and eight actual denoising iterations. Its only teacher targets were the pinned original FP32 Supertone Supertonic 3, CFG4, 50 steps. The original 50-step trajectory is interpolated onto the student's eight-step time grid; the final target is the exact original final latent. The student is a learned approximation, not mathematically identical to CFG4 or the original model.

The F1200 vector SHA256 is 7c81ab7f0ddcc4266743701599a8708d2edd1be04285770f996e84a8bb5f4447. Source revision: 13fb103. Source and experiment records: GitHub project.

Optional2x package: F1200 with internal latent retiming

models/round8-latent-retime-g135/ preserves all F1200 learned weights. It runs eight actual denoising calls at internal1.35 duration and resizes24-channel fine latent frames inside the original vocoder to produce2x speech. No waveform tempo/pitch filter or extra learned head is used. Other requested speeds retain the native duration path. Default F1200 remains available unchanged.

Use the project runtime, which supplies the new target_frames vocoder input. The GitHub downloader pins this package's immutable revision and verifies file hashes.

python download_artifacts.py --purpose inference --model round8-latent-retime-g135
python synthesize.py '안녕하세요. 오래된 컴퓨터에서도 빠르게 음성을 생성합니다.' \
  --model .hf-assets/model/models/round8-latent-retime-g135 \
  --lang ko --voice M1 --steps 8 --speed 2 --output samples/fast.wav

The user confirmed less mechanical noise and acceptable pronunciation on a short male Korean A/B sample. This is limited listening feedback. On sixteen consumed validation contexts, KO CER is51/193(26.42%) and EN WER1/96(1.04%), versus native2x F1200's133/193 and68/96. Normal1.05 F1200 is5/193 and0/96. Substantial Korean errors remain. This is not an omission-free general model, independent test, MOS, or speaker-equivalence result.

Optional enginev4 on the shared EPYC7551 one-logical-CPU/1GB server has paired median speedups1.042x KO and1.047x EN over enginev2 with unchanged waveforms. Short2x RTF is about1.35, long2x1.32..1.33, normal1x about1.00, excluding loading/file I/O. There is no realtime guarantee or new ARM result. The package is standard ONNX; the optional Linux AVX2 kernel must be built separately from GitHub source and is not shipped here.

This graph transform needs no new training dataset or checkpoint. Existing original50 reproduction data and F1200 reproduction checkpoints remain unchanged. All package file hashes are added to asset_manifest.json; the upstream license is preserved. The engine and preparation source are recorded at GitHub commit 7704219; see its9th-round record for validation and failed trials.

Models retained

• models/round5-original50-student/: selected F1200, 1,200 training updates; vector file48,903,928bytes.
• models/round5-original50-preview/: frozen E400 for listening comparison; not the default. Same architecture as F1200.
• models/cfg-mixed-v1-selected/: historical B baseline, retained for comparisons.
• models/cfg1024-v1-final/, models/mixed-v1-selected/: historical published experiments. Their quality/performance records do not apply to F1200.

Rejected derived-teacher training artifacts and the C/D model rejected for smearing/mechanical sound are not part of this release. No model pruning or retraining is planned after this F1200 selection; subsequent work optimizes the fixed-model engine.

Measured scope

On the shared EPYC7551 one-core, approximately1GB host, three alternating E/F process pairs gave F1200 medians5.594s/RTF1.0515 Korean and5.280s/RTF1.0702 English for the fixed short sentences. Loading and file writes are excluded; CPU affinity and ORT/BLAS threads are fixed to one. The sampled50ms memory monitor is not a cgroup limit. These measurements do not predict ARM speed. The historical ARM PR measures B, not F1200.

A four-sentence development pilot (KO2/EN2) had0/44 Korean character errors and0/20 English word errors for original50, B, E400 and F1200. It is too small for broad quality claims, and no fresh independent F1200 test has been evaluated. Existing test sets have already been consumed.

Minimal reproduction assets

reproduction/round5/original50-initial.pt and original50-step-01200.pt preserve the exact F1200 initialization and final checkpoint. The directory also contains training and initialization reports, corpus and a teacher-data audit. Intermediate checkpoints200–1000 and rejected workflows are excluded. E400 is supplied as an inference comparison package; its separate training checkpoint is omitted to limit size.

The companion dataset contains640train/128validation pairs from80/16 unique authored sentences, generated only by the original50-step teacher. No test trajectories were collected. The initialization reuses existing student weights as an initial value, never as teacher targets. Six replacement FFNs were previously fit only to original FP32 teacher activations. The uploaded initial checkpoint means those fits do not need to be repeated to reproduce this training stage.

See the source reproduction instructions. New training is not guaranteed bit-identical across hardware/library versions. asset_manifest.json lists SHA256/byte length for every artifact. Public downloads require no token. Source code has a separate license; all derivative model weights retain the upstream Open RAIL-M terms.

Final exported runtimes — project closed on 2026-11-02

Development stopped at GitHub source commit d9340482a2626dfa3d9c04b684369082674aa8f3. The F1200 learned weights and quantization are unchanged. The complete final exported ONNX packages are preserved in addition to the portable source models:

• models/F1200-retime-v5/: Linux x86_64 AVX2 engine v5, including the prepared batch-one graph. Build the optional native integer kernel from the GitHub source on a compatible host.
• models/F1200-retime-arm64/: the final all-parts row-normalization package tested on Mac M4 Max. Build native/build_row_norm.py from the GitHub source. This is not an Oracle ARM benchmark or a precompiled binary.

Both contain the voice styles, configuration, license and all ONNX graphs. Eight actual denoising steps are required. They are architecture-specific exports, not smaller learned models. The portable round5-original50-student and round8-latent-retime-g135 packages remain available.

The final shared 1-core/~1GB x86 server also runs other processes; repeat timings and version rankings are inconsistent. See the complete raw measurements and limitations at https://github.com/8954sood/supertonic3-optimization/blob/main/records/2026-10-30/ROUND18_REMEASUREMENT.md. Do not interpret one RTF value or a small difference as a performance guarantee. Short listening checks do not establish general pronunciation equivalence at high speeds.

The original50 initial/final checkpoints remain under reproduction/round5/. The training dataset remains pinned at babihoba/supertonic3-optimization-data@e56d93c9959647c850b639a614e629e4e4ac1593. No new training data or learned weights were created for these runtime exports. Full source and restoration instructions: https://github.com/8954sood/supertonic3-optimization.

A SHA-256 inventory of the final exports is stored at releases/2026-11-02/final_exports_manifest.json.