Fix silent all-zero fbank in chunk embedding prep when the 30s model is absent

[attevon-admin]

Closes #16

The bug

ChunkPrep::compute_chunk_fbank (src/pipeline/chunk_embedding/prep.rs) nested the "does this chunk fit the 30s model?" test inside the "is the 30s model loaded?" test:

let mut fbank = vec![0.0f32; self.largest_fbank_frames * 80];

if chunk_audio.len() <= 480_000 {
    if let Some(fbank_model) = &self.fbank_30s { /* 30s path */ }
    // fbank_30s == None falls through with `fbank` still all zeros
} else if let Some(fbank_model) = &self.fbank_10s { /* 10s path */ }

Ok(fbank)

Because the 10s branch is the else if of the length check, it is unreachable for chunks <= 480_000 samples. With fbank_30s == None a short chunk therefore returns the zeroed buffer allocated up front — a well-formed, all-zero tensor. No error, no panic; just silently wrong embeddings downstream.

run_sequential_chunks in orchestrate.rs:240-269 already does this correctly, keying the fallback off whether the 30s fbank was actually produced rather than off chunk length.

The fix

Extracted the choice into a small pure function so the two conditions are combined and the decision is explicit and testable:

pub(super) fn select_fbank_path(chunk_len: usize, has_30s: bool, has_10s: bool) -> FbankPath {
    if chunk_len <= 480_000 && has_30s {
        FbankPath::ThirtySecond
    } else if has_10s {
        FbankPath::TenSecond
    } else {
        FbankPath::None
    }
}

compute_chunk_fbank now dispatches on FbankPath, so a short chunk with no 30s model falls back to the 10s path exactly as orchestrate.rs does. The 30s and 10s computation bodies are unchanged — this is purely a control-flow correction.

Tests

Four unit tests in prep.rs cover the matrix (30s available, 30s missing, long chunk, no models). The regression test is short_chunk_falls_back_to_10s_when_30s_missing.

Verified on real Apple Silicon hardware (arm64 Mac Studio), since chunk_embedding is gated behind #[cfg(feature = "coreml")] and does not compile on Linux (the objc2* dependencies are Apple-only):

cargo test --release --no-default-features --features openblas-system,online,coreml

• Before the fix (original nested logic restored, same machine), the regression test fails precisely as expected while the other three pass:

  test ...prep::tests::short_chunk_falls_back_to_10s_when_30s_missing ... FAILED
  assertion `left == right` failed
    left: None
   right: TenSecond
  

  None is exactly the state that yielded the all-zero fbank.

• After the fix, all four pass and the lib suite is 88 passed; 2 failed out of 90.

The 2 failures (fast_apple_embeddings_match_python_fixture, fast_apple_split_primary_batch_matches_single_tail_path) are pre-existing and unrelated — I confirmed they fail identically with stock, unmodified prep.rs on the same machine (fixture drift in my local model artifacts, not a regression from this change).

Summary by CodeRabbit

• Bug Fixes
  • Improved audio feature model selection for short audio segments.
  • Short segments now use the 10-second model when the 30-second model is unavailable, instead of producing empty results.
• Tests
  • Added coverage for model selection across different segment lengths and model availability conditions.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 7f2a56de-f479-47cf-89e6-37dc6d4c69df

📥 Commits

Reviewing files that changed from the base of the PR and between b0756b1 and fe7164b.

📒 Files selected for processing (1)

• src/pipeline/chunk_embedding/prep.rs

Included review availability: Your plan provides up to 2 included reviews per hour; 1 remains after this review.

---

📝 Walkthrough

Walkthrough

The chunk embedding preparation path now selects an available fbank model. Short chunks fall back to the 10-second model when the 30-second model is unavailable. Unit tests cover selection and no-model cases.

Changes

Fbank model selection

Layer / File(s)	Summary
Fbank selection and preparation src/pipeline/chunk_embedding/prep.rs	FbankPath and select_fbank_path select the 30-second model, the 10-second model, or no model. compute_chunk_fbank uses this selection before prediction. Unit tests cover short and long chunks, fallback behavior, and unavailable models.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to fe716

This localized change makes short chunks fall back to the available 10-second model instead of producing an all-zero feature buffer; no actionable merge-blocking risk remains after normal checks and review.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 71.43% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 7 functions across 1 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the primary change: preventing silent all-zero fbank output when the 30-second model is absent.
Linked Issues check	✅ Passed	The changes address issue #16 by selecting the 10-second model when a short chunk lacks the 30-second model. The added selection logic and regression tests support the required fallback behavior.
Out of Scope Changes check	✅ Passed	The changes are limited to fbank model selection and related tests in chunk embedding preparation. No unrelated changes are identified.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Greptile Summary

The PR corrects Core ML chunk feature-bank routing so short chunks fall back to the 10-second model when the 30-second model is unavailable.

• Introduces an explicit, testable feature-bank path selector.
• Preserves the 30-second path for eligible chunks and the 10-second path for longer chunks.
• Adds unit coverage for model-availability and chunk-length combinations.

Confidence Score: 5/5

The PR appears safe to merge, with no actionable regressions identified in the corrected feature-bank routing.

The selector sends eligible chunks to the 30-second model and otherwise uses the available 10-second model, matching the established sequential implementation without introducing new buffer, shape, or availability failures.

Important Files Changed

Filename	Overview
src/pipeline/chunk_embedding/prep.rs	The new selector fixes the silent all-zero fallback while preserving existing model limits, buffer bounds, and sequential-path semantics.

_{Reviews (1): Last reviewed commit: "fix(chunk-embedding): fall back to 10s f..." | Re-trigger Greptile}