[chaofengw-nv]

Background

Fixes #1749. Approximate audio resampling changes Mimi input codes and can redirect PersonaPlex generation. The bounded Kaiser filter in #1751's earlier revision 8f6eddc02 changes clip 18's codes at frame 23: the same engine passes the original speech gate with Native's resampled input (1.08 dB) and fails with the Kaiser input (3.78 dB).

Exit Criteria

• Follow Native's HQ resampling recipe and librosa output length, including zero padding.
• Keep the dependency within PersonaPlex without requiring development headers or a link-time libsoxr dependency for every model family.
• Preserve the original requests, Native baseline, FP16 temporal/depth core and 3 dB speech gate.

Implementation

Load the system libsoxr.so.0 when resampling is needed. Its opaque C API uses soxr_create defaults (HQ, float32 I/O, one thread), consumes the input and flushes until drained or the output is full. Already-matching sample rates copy the input. A missing library produces an actionable error.

The family CI dependency image installs libsoxr0; the family README documents this separate LGPL-2.1-or-later runtime dependency. The wheel does not vendor it. There is no global CMake dependency search or global image change.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

• python -m pytest -c /dev/null -o addopts= -p no:cacheprovider -q families/personaplex/tests/test_resampler_precision.py: same final test on both CPU controls. Kaiser revision 8f6eddc02 (before splitting fix(personaplex): align FP16 inference with Native #1751): 1 numerical failure, 5 passes, zero errors/skips. Final implementation: 6 passes, zero errors/skips. Local execution without libsoxr: identity/invalid-rate/missing-library assertions pass; 5 waveform cases explicitly skip.
• Original 20 input waveforms: output lengths match Native; maximum absolute error is 7.152557e-7, not bit-exact. Final flush-loop outputs are bit-identical to the GPU-tested prototype on all 20 clips.
• One-engine injection control: Native waveform passes clip 18 at 1.08 dB; Kaiser waveform fails at 3.78 dB and reproduces the earlier output digest. The engine SHA is identical in both runs.
• Final published-code integration: 18/20, with clip 18 passing. Clip 7 fails at 5.58 dB and clip 11 at 6.42 dB under the unchanged gate, against both frozen and fresh same-batch Native FP16. Fresh Native PCM and output-frame counts match the frozen reference on all 20 requests. Exact input re-exposes clip 7, which had passed with altered Kaiser input codes. This is not whole-model acceptance.
• Final same-batch Task P50: Native 2125.46 ms / TRTMC 1474.17 ms, 20 requests each; output sample counts and rates match on all requests (4,617,600 samples per side). The clean integration passes 7 family CTests, 6 resampler CPU tests and 7 RMSNorm-gain GPU tests. readelf -d confirms no libsoxr DT_NEEDED entry.
• ruff check, clang-format --dry-run --Werror, git diff --check, python -m tools.model_ci validate, and python tools/test_impact.py --validate: passed.

Hardware, Environment, and Revisions

Base 9a8b545f5e8c22c653c8029c95a999dc070b4f3f; head b3b12608132d5935c264f4c0615e1dce25e1e470. GB300, Linux ARM64, CUDA 13, TensorRT 11.2.0.113, PyTorch 2.12.0+cu130. Native uses librosa 0.11.0, python-soxr 1.1.0 and bundled libsoxr 0.1.3-14-ga66f3ee. Checkpoint nvidia/personaplex-7b-v1@fdaf4090a61cb315c138a1faee287ffd6c716309; original 20 LibriSpeech requests.

Final runtime source SHA 723604cd...; final test SHA 583ad9a5.... Final GPU integration uses this runtime from a clean composition of the published PR heads, including #1751 decoder-only and #1800 gain-only. Its 1,528 production Python/CLI files match the builder source exactly, so the same engine bundle is explicitly reused by SHA. The temporal engine and all 16 depth engines remain Half; Mimi remains FP32.

Not Run / Remaining Gaps

Three-build model acceptance, current-head premerge, a rebuilt installed wheel and other hardware are unqualified. The system libsoxr runtime must be installed for resampling. The Native codec-policy decision and clip-7 core-numerics investigation remain separate; neither the Native baseline nor the speech tolerance was changed.

Contributor Self-Review

• I have completed a self-review of this change.

Reviewed the C API signatures/defaults, loader lifetime, flush, length bounds, all PersonaPlex callers, dependency ownership and the same-test original/fixed controls. Independent review found no blockers after the flush and missing-library test improvements.

Notes For Future Readers

The PR scopes are now split: #1751 (head 8ae0f442) carries the Mimi FP32 decoder fix for #1750; this PR carries the resampler fix for #1749. They can merge independently. This branch starts from current main and contains no decoder-precision change. The GPU integration above used both fixes alongside the other pending PersonaPlex arithmetic fixes.

The public libsoxr header documents soxr_create's HQ defaults. soxr_oneshot defaults to LQ, so substituting it without an explicit HQ specification changes the signal.

Risk level

• Low
• Medium
• High

Resampling changes conditioning for non-24 kHz audio and adds a family runtime dependency. Full PersonaPlex speech parity still fails on two requests.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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

[chaofengw-nv]

Superseded by #1751. This change and its regression coverage are preserved in commit 258c706 of the consolidated PR. Original issue links and acceptance criteria remain tracked there.