Pin librosa <0.11 to protect voicing-gate calibration; fix fork HQ lane test stub by KhryptorGraphics · Pull Request #2 · KhryptorGraphics/AutoVoice

KhryptorGraphics · 2026-09-07T02:08:30Z

Two fixes found while auditing the multi-speaker conversion path.

1. Pin `librosa>=0.10,<0.11` (`762f0b7`)

librosa 0.11 changed pyin's voiced decision: white noise scores ~0.43 voiced fraction on 0.11 vs 0.0 on 0.10.2 (probed on both versions). Every multi-speaker routing gate — multi_speaker_merge_voiced_min and friends, calibrated on leads 0.76–0.87 vs keep-cases 0.14–0.51 — was measured on 0.10 semantics. With the previous open bound (librosa>=0.10), any fresh install, CI run, or redeploy resolving to 0.11 gets a silently mis-calibrated router and a failing test_multi_speaker_routing noise test.

requirements-runtime.txt: bound to >=0.10,<0.11 with a comment saying re-calibrate before raising it
_voiced_fraction docstring: notes the calibration coupling
requirements.lock: refreshed sha256 + line_count (follows the checker's runtime_requirement_count, now 32); scripts/check_dependency_contract.py passes with no errors

2. Fix `test_fork_hq_lane` stub signature (`a4b67eb`)

The quality-tiers change (8ad0500) made _convert_song_fork_hq call _extract_pitch(..., method=q['f0_method']), but the test's monkeypatched lambda kept the old (audio, sr) signature, so test_fork_hq_lane_is_stereo_and_mirrors_contract failed at HEAD with a TypeError. The stub now mirrors the real signature's optional kwarg.

Verification

test_separation_bridge + test_multi_speaker_routing + test_svc_fork_bridge: 71/71 pass under librosa 0.10.2
test_fork_hq_lane + test_web_conversion_metadata + test_conversion_workflow_api: 16/16 pass
scripts/check_dependency_contract.py: no errors
cd frontend && npm run build: clean
Not runnable in the audit container (expected to pass on the Jetson): 14 pipeline tests needing the HiFiGAN checkpoint, 4 platform-script tests needing the autovoice-thor conda env

🤖 Generated with claude-flow

https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

Generated by Claude Code

Summary by Sourcery

Stabilize multi-speaker voicing calibration and fork HQ coverage while making CI and live test workflows portable to CPU-only environments.

Bug Fixes:

Pin librosa below 0.11 to preserve the calibrated multi-speaker voicing-gate behavior.
Fix the fork HQ lane test stub to support the pitch-extraction method argument.
Prevent CUDA cache operations from running when CUDA is unavailable.
Ensure karaoke uploads create their temporary directory before creating the temporary file.

Enhancements:

Make live frontend tests and the deterministic backend harness run reliably across Jetson and CPU-only environments.
Use deterministic SQLite, GPU, and benchmark fixtures for live backend UI testing.
Skip CUDA-dependent coverage and contract tests on machines without GPU support.

CI:

Install CPU-only PyTorch in CI to avoid pulling large NVIDIA wheel dependencies.
Exclude CUDA-marked tests from CPU-only CI contract and coverage runs.
Make Playwright select the Jetson Python environment when available and fall back to python3 elsewhere.

Documentation:

Document the dependency between voicing-gate calibration and librosa 0.10 semantics.

Tests:

Update live workflow tests for explicit generation-start interaction.
Make training, platform, and live backend tests deterministic on CPU-only CI environments.

Chores:

Refresh the frontend dependency lockfile and runtime dependency lock metadata.

librosa 0.11 changed pyin's voiced decision: white noise scores ~0.43 voiced fraction vs 0.0 on 0.10.x (probed on both versions). Every multi-speaker routing gate (multi_speaker_merge_voiced_min and friends, calibrated leads 0.76-0.87 vs keep-cases 0.14-0.51) was calibrated on 0.10 semantics, so an environment resolving librosa>=0.10 to 0.11 gets a silently mis-calibrated router (and a failing test_multi_speaker_routing noise test). Bound the requirement to the calibrated major, note the coupling in _voiced_fraction, and refresh the requirements-runtime hash in the dependency contract lock (line_count follows the checker's runtime_requirement_count, now 32). Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

The quality-tiers change made _convert_song_fork_hq pass method=q['f0_method'] to _extract_pitch, but the test's monkeypatched lambda still had the old (audio, sr) signature, so test_fork_hq_lane_is_stereo_and_mirrors_contract failed at HEAD with a TypeError. Mirror the real signature's optional kwarg in the stub. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-09-07T02:08:33Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

chatgpt-codex-connector · 2026-09-07T02:08:36Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

sourcery-ai · 2026-09-07T02:08:37Z

Reviewer's guide (collapsed on small PRs)

Reviewer's Guide

The PR protects multi-speaker routing behavior by pinning librosa to the pyin semantics used during calibration, documenting and locking that dependency contract, and fixes the fork HQ lane test stub to accept the new pitch-extraction keyword argument.

Sequence diagram for the fork HQ lane pitch extraction test

sequenceDiagram
    participant T as test_fork_hq_lane_is_stereo_and_mirrors_contract
    participant C as _convert_song_fork_hq
    participant E as _extract_pitch
    T->>C: invoke fork HQ conversion
    C->>E: _extract_pitch(audio, sr, method=q['f0_method'])
    E-->>C: pitch result
    C-->>T: stereo conversion result

Flow diagram for librosa calibration protection

flowchart LR
    R["requirements-runtime.txt: librosa>=0.10,<0.11"] --> P["librosa 0.10 pyin semantics"]
    P --> V["_voiced_fraction"]
    V --> G["Multi-speaker routing gates"]
    U["librosa 0.11"] --> W["Changed voiced decision"]
    W --> M["Mis-calibrated routing"]

File-Level Changes

Change	Details	Files
Constrain the runtime librosa version to preserve the voicing-gate calibration contract.	Pin librosa to >=0.10,<0.11 and document that recalibration is required before upgrading. Refresh the locked dependency metadata and update the voiced-fraction documentation with the librosa-version coupling.	`requirements-runtime.txt` `requirements.lock` `src/auto_voice/inference/singing_conversion_pipeline.py`
Align the fork HQ lane test double with the production pitch-extraction call signature.	Allow the monkeypatched extractor to accept the optional f0 method keyword argument.	`tests/test_fork_hq_lane.py`

Tips and commands

Interacting with Sourcery

Trigger a new review: Comment @sourcery-ai review on the pull request.
Continue discussions: Reply directly to Sourcery's review comments.
Generate a GitHub issue from a review comment: Ask Sourcery to create an
issue from a review comment by replying to it. You can also reply to a
review comment with @sourcery-ai issue to create an issue from it.
Generate a pull request title: Write @sourcery-ai anywhere in the pull
request title to generate a title at any time. You can also comment
@sourcery-ai title on the pull request to (re-)generate the title at any time.
Generate a pull request summary: Write @sourcery-ai summary anywhere in
the pull request body to generate a PR summary at any time exactly where you
want it. You can also comment @sourcery-ai summary on the pull request to
(re-)generate the summary at any time.
Generate reviewer's guide: Comment @sourcery-ai guide on the pull
request to (re-)generate the reviewer's guide at any time.
Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
pull request to resolve all Sourcery comments. Useful if you've already
addressed all the comments and don't want to see them anymore.
Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
request to dismiss all existing Sourcery reviews. Especially useful if you
want to start fresh with a new review - don't forget to comment
@sourcery-ai review to trigger a new review!

Customizing Your Experience

Access your dashboard to:

Enable or disable review features such as the Sourcery-generated pull request
summary, the reviewer's guide, and others.
Change the review language.
Add, remove or edit custom review instructions.
Adjust other review settings.

Getting Help

Contact our support team for questions or feedback.
Visit our documentation for detailed guides and information.
Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.

sourcery-ai

Hey - I've reviewed your changes and they look great!

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

greptile-apps · 2026-09-07T02:11:06Z

Greptile Summary

The PR protects calibrated multi-speaker voicing gates by constraining librosa to its 0.10 behavior and updates the HQ-lane test stub for the current pitch-extraction signature.

Pins librosa below 0.11 and refreshes dependency-contract metadata.
Documents the coupling between _voiced_fraction calibration and librosa semantics.
Allows the HQ-lane test double to receive the pitch-extraction method, although the forwarding contract remains unasserted.

Confidence Score: 4/5

The PR appears safe to merge, with one non-blocking test-strengthening opportunity around HQ pitch-method forwarding.

The dependency constraint is consistently integrated and the production behavior is unchanged; the only accepted concern is that the revised stub can conceal a future loss of the f0_method argument.

Files Needing Attention: tests/test_fork_hq_lane.py

Important Files Changed

Filename	Overview
requirements-runtime.txt	Constrains librosa to the calibrated 0.10 release series with an explanatory comment.
requirements.lock	Refreshes the runtime requirements hash and count consistently with the manifest.
src/auto_voice/inference/singing_conversion_pipeline.py	Documents the version-sensitive pyin calibration without changing runtime behavior.
tests/test_fork_hq_lane.py	Fixes the test-double signature but does not verify that a configured non-default pitch method is forwarded.

_{Reviews (1): Last reviewed commit: "fix: accept method kwarg in fork HQ lane..." | Re-trigger Greptile}

greptile-apps · 2026-09-07T02:11:09Z

+    monkeypatch.setattr(p, "_extract_pitch",
+                        lambda a, sr, method=None: np.zeros(8, np.float32))


Pitch forwarding remains untested

The updated stub accepts but ignores method, and this test does not configure an f0_method. It will therefore still pass if the HQ lane stops forwarding a selected pitch-extraction method. Exercise a non-None override and assert the received argument so this routing contract is protected.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

add_sample_from_path gated torch.cuda.empty_cache()/synchronize() on torch being importable, not on CUDA being present, so the endpoint 500'd on any CPU-only host (hosted CI runners included). Mirror the guard api_youtube already uses. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

Hosted CI has failed on every main push back to run 61; all five failing jobs traced to GPU/Jetson assumptions leaking into lanes that run on ubuntu-latest: - backend lane: deselect @pytest.mark.cuda tests (-m 'not cuda') — the marker exists for exactly this split - coverage gate: deselect cuda-marked tests only when torch.cuda is unavailable, so Jetson coverage numbers stay canonical - test_training_job_manager: stub torch.cuda.is_available in the fixture and init tests (queue/persistence logic under test; the GPU gate itself is covered by test_gpu_enforcement) - phase5 orchestration tests: fake auto_voice.models.diffusion_decoder — the default architecture constructed the real decoder on cuda - test_cuda_device_auto_selection: mark cuda (Trainer moves internal modules to the selected device, so stubbing is_available is not enough) - test_common_env_prefers_canonical_python: fake HOME with a stub canonical env python, deterministic on machines without autovoice-thor - preflight test: pass --no-require-gitnexus (gitnexus presence is orthogonal to the benchmark-report behavior under test) - live_backend_server harness: present a CPU-safe fake GPU so the production /ready probe (CUDA-gated by design) reports ready — fixes deployment-smoke's validate_release_candidate step - playwright.live.config.ts: fall back to python3 when the Jetson interpreter path does not exist — fixes browser-live's backend webServer startup Verified on a CPU-only container: backend lane 102 passed, previously failing suites 169 passed, /ready 200 + validate_release_candidate rc=0. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

npm audit fix within existing ranges; production audit now reports no high/critical findings (2 moderates remain in react-router, which needs a breaking major). Refresh the frontend lock hash in the dependency contract; checker passes. Frontend typecheck and build verified. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:03:43Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

f74d4c7 replaced workflow auto-submit with the Start Generation button, but both live specs still waited for the workflow card right after the uploads, so they timed out before Convert Workflow Song ever rendered (the browser-live lane has been red since). Click the explicit start-generation-button first. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:06:35Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

…harness pip install -e .[dev] on ubuntu-latest now resolves PyPI torch with the multi-GB nvidia-*-cu13 wheel stack and dies with ENOSPC before tests run (seen on run 41670270628). Pre-install the CPU wheel so the resolver keeps it — the runners have no GPU to feed anyway. The deterministic harness also now serves /reports/benchmarks/latest (and release evidence) from a seeded fixture dir: the System page fetch 404'd otherwise, and the browser's automatic console error failed the zero-console-error live acceptance check. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:09:33Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

The speaker endpoints 500'd in the harness (and would on CI): the DB layer defaults to the production MySQL server, which only exists on the Jetson. Default AUTOVOICE_DB_TYPE=sqlite before auto_voice imports (db.schema reads it at import time) and keep the sqlite file inside the harness's throwaway data dir instead of a cwd-relative path. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:12:09Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

upload_song passed dir=/tmp/autovoice_karaoke to NamedTemporaryFile and only ran os.makedirs afterwards, so the first upload on a host where the directory does not yet exist (fresh CI runner, fresh boot) failed with FileNotFoundError -> 503. Seen as the karaoke/full-ui live-spec failures on run 41671275025. Create the directory first. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:15:17Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

The speaker endpoints still 500'd under the harness sqlite DB: a fresh file has no tables (no such table: speaker_clusters). Run init_database('sqlite') after pointing DATABASE_PATH at the throwaway data dir. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:17:06Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

The remaining-module gate holds src/auto_voice/web/persistence.py to 80% but no suite in its list exercised the store directly — the API suites only graze it, leaving it at 67.7% on hosted CI (the gate's first-ever run there; the threshold stage never ran before because the pytest phase always died earlier). Add round-trip unit tests for the CRUD surfaces (training/background jobs, presets, history, workflows, diarization, checkpoints, webhooks, loaded models, youtube history + purge, asset registry, karaoke sessions) and the normalizer edge branches; the suite alone measures persistence.py at 91%. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:30:07Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

run_remaining_module_coverage.py writes it into the repo root (CI uploads it as an artifact); it is generated output, not source. Co-Authored-By: claude-flow <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01Vqqhm9yzsrG7rgmiMrBjC1

cr-gpt · 2026-11-03T16:30:50Z

Seems you are using me but didn't get OPENAI_API_KEY seted in Variables/Secrets for this repo. you could follow readme for more information

claude and others added 2 commits September 7, 2026 01:38

sourcery-ai Bot reviewed Sep 7, 2026

View reviewed changes

greptile-apps Bot reviewed Sep 7, 2026

View reviewed changes

claude and others added 3 commits November 3, 2026 16:03

		monkeypatch.setattr(p, "_extract_pitch",
		lambda a, sr, method=None: np.zeros(8, np.float32))

Conversation

KhryptorGraphics commented Sep 7, 2026 • edited by sourcery-ai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

1. Pin librosa>=0.10,<0.11 (762f0b7)

2. Fix test_fork_hq_lane stub signature (a4b67eb)

Verification

Summary by Sourcery

Uh oh!

cr-gpt Bot commented Sep 7, 2026

Uh oh!

chatgpt-codex-connector Bot commented Sep 7, 2026

Uh oh!

sourcery-ai Bot commented Sep 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Reviewer's Guide

Sequence diagram for the fork HQ lane pitch extraction test

Flow diagram for librosa calibration protection

File-Level Changes

Interacting with Sourcery

Customizing Your Experience

Getting Help

Uh oh!

sourcery-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot commented Sep 7, 2026

Greptile Summary

Confidence Score: 4/5

Important Files Changed

Uh oh!

greptile-apps Bot Sep 7, 2026

Choose a reason for hiding this comment

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

cr-gpt Bot commented Nov 3, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

KhryptorGraphics commented Sep 7, 2026 •

edited by sourcery-ai Bot

Loading

1. Pin `librosa>=0.10,<0.11` (`762f0b7`)

2. Fix `test_fork_hq_lane` stub signature (`a4b67eb`)

sourcery-ai Bot commented Sep 7, 2026 •

edited

Loading