[jadenfix]

Restores the missing packaged configuration and repairs SER construction order, complete-config checkpoint persistence, CNN shape calculation and handcrafted-feature forward/backward execution. Includes existing PR2 policy/TTS fixes, PR3 audio QA dropout/PPO/terminal fixes and PR1 launcher changes without modifying those owner branches. WAV decoding now uses the declared SoundFile backend, works with TorchAudio2.9 without optional TorchCodec, and validates actual output; explicit synthetic WAV mode respects configured sample rate.

Actual scoped workflow: synthetic 220Hz/440Hz tones -> real mel/pitch/energy extraction -> CNN-LSTM-attention SER with three Adam updates -> complete-config checkpoint restore -> identical normalized predictions -> greedy policy action. Four new SER tests and real WAV/checkpoint tests bring the integrated suite to35 passing tests. Installed wheel SER workflow runs outside the checkout. Exact-head CPU GitHub workflow38333894251 passes dependency installation, the full suite and actual SER artifact workflow on35bf3538fccf7dc1a808b68b35826f67cd398812.

The model starts random and learns arbitrary fixture labels; this proves computation and changed parameters, not emotion accuracy or a live voice product. Live TTS/microphone, semantic dataset validation and on-policy PPO assumptions remain outside scope. Merge coordinator: PR3 is stacked on PR2; this branch incorporates them and PR1, so preserve support changes when consolidating.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
human-voice-ai	Error		Oct 13, 2026 1:51am UTC

[jadenfix]

Independent exact-head audit: HOLD

Pinned identities:

• current main 3b40f00612a51059e96df4e492e401128f5688ad, tree d838069418dfb1dfc125f303fa3cc5f569881cc2
• audited PR head 35bf3538fccf7dc1a808b68b35826f67cd398812, tree d24ff3647ffbb8dd74a05037cbd3535a084b4c35

The narrow provider-free computation is real and reproducible. On the exact head:

• PYTHONPATH=src HF_HUB_OFFLINE=1 python -m pytest -q -> 35 passed.
• The scoped human_voice_ai.ser_smoke run produced finite mel/pitch/energy features with shapes [65,18], a [2,1,16,65] mel batch, three Adam updates, nonzero parameter change and bit-identical probabilities after strict checkpoint restore. The receipt correctly says no emotion-accuracy, live-TTS or microphone claim.
• I built human_voice_ai-0.1.0-py3-none-any.whl, installed it into a separate target and ran the SER workflow from /tmp outside the checkout; packaged config import and the receipt passed.
• Exact-head Actions run 38333894251 is green for this same 35-test/synthetic-SER lane.

A core real-audio defect remains. Reproduction:

t = np.arange(4096, dtype=np.float32) / 16000
stereo = np.stack([
    .1 * np.sin(2*np.pi*220*t),
    .1 * np.sin(2*np.pi*440*t),
], axis=1)
sf.write("stereo.wav", stereo, 16000)
FeatureExtractor(sample_rate=16000, n_fft=128, hop_length=64,
                 win_length=128, n_mels=16).extract("stereo.wav")

soundfile.read returns (4096, 2) (frames × channels). extract() unconditionally adds a leading dimension, producing a 3-D tensor, and extract_from_waveform() raises:

ValueError: Expected 1D or 2D input, got 3D

This is not an exotic malformed input: a normal stereo WAV cannot enter the packaged SER flow, even though lower extractor stages explicitly contain multi-channel averaging logic and the PR describes actual WAV support.

Required before functional clearance:

1. normalize SoundFile output to the documented channels × time contract (or explicitly and truthfully downmix);
2. add direct mono/stereo WAV regressions, including resampling, finite output and expected shape/content checks;
3. rerun the full suite and installed-wheel workflow on the repaired exact head;
4. keep the current truth boundaries—synthetic labels do not establish emotion accuracy or a live voice product.

Release/demo gates also remain incomplete: Vercel status is failed, the broader matrix CI has no exact-head run in the current PR-triggered run set, and there is no source-pinned GIF/MP4, capture receipt or media-integrity verification. This therefore earns no demo/current-main/media coverage.

Ownership/dependency note: #4 is the canonical backend integration of #1/#2/#3. Draft #5 is a disjoint frontend-integrity delta still based on #3; it must be reconciled onto the eventually repaired integration rather than losing its “no fabricated inference” behavior. I did not modify, push, merge, close or retarget any branch.