Current optimized-pipeline validation
Five-language listening release
The public Podcast listening Space combines the existing six EN/DE examples with nine newly generated NL/FR/ES episodes. The extension produced 420 candidates and 180 selected turns at N=4/2/1 on one shared RTX 3090, using the same reference pair and sparse prompting. Strict ranking, correct reference routing, exact candidate-to-selected PCM, transcript/prompt preservation, MP3 hashes and decoding, estimated-boundary joins, and timing accounting passed artifact validation for all three modes. Sixteen CPU tests pass.
All 17 players (15 episodes and two references) progressed in Chromium both locally and on the public HTTPS Space; candidate order, language navigation, transcripts, prompt-copy feedback, exclusive playback, and 1440/768/390-pixel layouts passed. These playback checks do not imply perceptual correctness of every spoken word.
The extension has 49 configured review flags among 180 selected turns: 39 transcript checks exceed 15% ASR WER, and 38 have uncertain boundaries, with overlap between those sets. Dutch contributes many boundary flags because it uses conservative Parakeet TDT estimates rather than unsupported Qwen forced alignment. Uncertain clips stay untrimmed. French and Dutch have possible omissions or language drift; ASR can also be wrong. Spanish has no review flags at N4/N2 and one at N1. Nothing was silently regenerated or reselected. See aggregate evidence and all per-turn records.
Earlier EN/DE runs
The new default was executed end to end on one RTX3090:40turns,160candidates, both DE/EN, exact prefill and codec/token/PCM checks, independent audit and MP3 export. All seven players in the listening comparison progressed in Chromium on desktop/mobile. English turn6 had a real incomplete top-ranked take; strict output is retained and a separately documented complete copy uses another candidate. See measurements and limits and the benchmark receipt.
Fourteen CPU tests cover the original cache/reward contracts, sparse cue word preservation, failed archival receipts, new default selection and custom script validation. Run cd optimized_pipeline && python -m unittest test_contracts test_speaker_recipe test_spontaneous test_inputs.
The listener has approved the spontaneous samples; that is not a blinded study or a quality guarantee. The fresh N=2/N=1 comparison is complete: 120 candidates, 80 selected turns, four MP3 episodes; zero configured content/alignment review flags. All seven listening-comparison players actually played in Chromium; desktop/mobile checks passed. Exact tested prefill/token and fused-decoder checks passed. See aggregate comparison and per-turn records. Native pairs, optional VC and internal pause compaction belong to the preserved legacy route; the prior evidence follows.
Validation status
This page separates the trained model evidence from newly packaged tooling.
- The source full checkpoint at update 9,874 completed training successfully and was used for 1,536 generated/scored candidates in the controlled EN/DE segment/block experiment. All 160 original episode conditions assembled, and 271 selected pause variants received full-dialogue evaluation.
- The BF16 deployment export was checked tensor by tensor against the source FP32 checkpoint cast to BF16, matching the evaluated loading precision. Shard checksums are in EXPORT_PROOF.json.
- The prior CPU-only script-boundary edit produced 1,416 trimmed previews and 123 complete montage conditions. 120 candidate boundaries required review; 37 montage conditions therefore had no safe strict-endpoint edited version. All original audio remained available.
- The prior local embedded MP3 and public players passed real Chromium playback, exclusive-player, trimmed/raw-toggle and responsive-layout checks.
- New portable package CPU tests cover script-vs-extra-ASR boundaries, internal hallucination rejection, waveform-word preservation, exact180-ms quiet-gap compaction, non-quiet-gap preservation, no constant-background crossfade gain bump, exact400-ms joins, global timestamp handling, identity-collapse penalty, prompt routing/timing, heterogeneous batch routing with tiny CPU fakes, CFG1 behavior, and target-only twelve-codebook training labels.
- The new internal-pause compaction remains UNVALIDATED for listening quality despite its signal-level tests.
- The optional split/convert/re-ASR/reassemble workflow is experimental and not validated end to end in the published listening grid.
- The fresh allocated four-GPU functional smoke passed for EN/DE, turns/pairs, two candidates, identity scoring, ASR, assembly and the quiet-pause option. It does not establish perceptual quality or validate optional VC.
- The historical trainer completed on the original 32-GPU environment. The new portable continued-training entry point is separate release tooling; its packer has CPU contract tests, not a completed fresh multi-GPU retraining result.
Reproduce the CPU checks:
python -m unittest discover -s tests -v
The exact release-time test report is provided in stats/release_validation.json. It must be interpreted together with the unvalidated-feature labels above. Automatic scores do not replace a listening study.
The successful v2 smoke generated 16 candidates;10 passed the configured content checks and8 selected units formed four complete episodes (EN turns11.35s,EN pairs19.51s,DE turns11.99s,DE pairs20.07s). The quiet-pause option was enabled, but no real pause met its conservative trigger in these short examples; actual splice behavior is covered only by synthetic CPU tests and remains perceptually unvalidated. Detailed result · The four new audio outputs.