A WAV or FLAC file that decodes without throwing an exception can still be a catastrophic training example. The defects that degrade deep learning models the most are precisely the ones a human ear misses during a quick spot-listen: brickwall spectral cutoffs from fake upsampling, destructive phase cancellation during mono downmixing, label drift in timed transcripts, and sub-segment train/validation leakage.
This article introduces an open-source, reproducible pre-training quality-assurance framework developed at ComposerMind Labs. By evaluating audio corpora across signal forensics, partition isolation, and annotation alignment before compute cycles are spent, this framework provides automated, deterministic quality gates for machine learning workflows.
+-------------------------------------------------------------------------+
| COMPOSERMIND AUDIOINTEGRITY WORKFLOW |
| |
| Raw Audio Corpus (Web Crawls, Multi-Track Stems, Transcripts) |
| | |
| v |
| [ STAGE 1: SIGNAL FORENSICS ] ---------------------------------------+ |
| * WaveWash (v0.2.0) --> Clipping, Saturation, DC-Offset Bias | |
| * BandCheck (v0.1.1) --> Fake 48kHz Upsampling, Spectral Cutoffs | |
| * PhaseCheck (v0.1.0) --> Inter-Channel Mono Cancellation Loss | |
| * DriftCheck (v0.1.0) --> Timed Label vs. Audio Offset, Drift (ppm) | |
| | | |
| v | |
| [ STAGE 2: PARTITION SECURITY ] -------------------------------------+ |
| * SplitSafe (v0.1.1) --> Acoustic Fingerprint Leakage Firewall | |
| | | |
| v | |
| [ STAGE 3: ANNOTATION INTEGRITY ] -----------------------------------+ |
| * TonicFind (v0.1.0) --> 12D Chromagram vs. Musical Key Profiles | |
| * TempoCheck (v0.1.0) --> Onset Envelope vs. BPM Octave Ambiguity | |
| | | |
| v | |
| Verified Audit Manifest --> Triage & Review --> PyTorch / HF Loader |
+-------------------------------------------------------------------------+
—————————————————————————————————————————————
- Why Audio Datasets Require Pre-Training Audits In natural language processing and computer vision, automated deduplication, normalization, and syntax validation are standard pre-requisites. In audio machine learning, however, ingestion pipelines often verify only that soundfile.read() or torchaudio.load() successfully loads an array. This surface-level validation permits silent, high-impact failure modes to contaminate training corpora: Failure Mode A: Sub-Segment Validation Leakage Consider a recording placed in the training partition, while a slightly trimmed (e.g., 50 ms offset), gain-adjusted, or re-encoded copy sits in the validation partition. Standard hash-based deduplication (MD5/SHA-256) fails completely because compression framing and trimming alter bit-level signatures. The model memorizes the underlying acoustic features, validation loss can plummet deceptively, and the model subsequently underperforms on genuine out-of-distribution audio. Original Audio (Train Set): =================================== Trimmed Copy (Validation Set): ============================= Result: File-level hashing misses the duplicate -> Silent cross-split contamination occurs.
Failure Mode B: Destructive Mono Downmixing
Many speech and audio pipelines downmix stereo inputs to mono (M = (L + R) / 2) prior to computing mel-spectrograms or feeding audio into acoustic backbones. When a stereo stem or recording contains inverted channel polarity (R ≈ -L) due to microphone miswiring or stereo widening effects:
M = (L + (-L)) / (2) ≈ 0
The waveform collapses toward digital silence. Even with partial phase cancellation, destructive interference selectively notches out specific frequency bands, training neural models on acoustically corrupted spectra.
Left Channel (L): /\ /\ /\ /\ (Active audio signal)
Right Channel (-L): \/ \/ \/ \/ (Inverted polarity)
-------------------
Mono Sum (L + R): ------------------- (CANCELLATION: near-total signal null)
Failure Mode C: Codec Cutoffs in High-Sample-Rate Containers
Web-crawled audio is frequently transcoded from legacy lossy formats (such as 128 kbps MP3) into uncompressed 48 kHz 24-bit WAV containers. While the container header claims a 24 kHz Nyquist bandwidth, signal energy abruptly drops off at 11 kHz or 16 kHz. When generative diffusion models or neural codecs are trained on such files, the loss function penalizes absent high frequencies, causing the model to risk learning container-induced spectral artifacts and unnatural cutoffs.
Failure Mode D: Timed Label Misalignment and Clock Drift
In speech recognition and audio-captioning datasets, audio files are paired with timed transcript labels (e.g., SRT, WebVTT, CSV, or Lhotse manifests). Clock drift between capture hardware or frame-rate mismatches (such as 25 fps subtitles mapped onto 23.976 fps video audio) cause annotations to drift systematically over time, leading the acoustic model to associate phonetic tokens with incorrect acoustic frames.
—————————————————————————————————————————————
- Taxonomy of Dataset Defects The ComposerMind audit framework categorizes pre-training defects into three operational domains: +-------------------------------------------------------------------+ | AUDIO CORPUS | +-------------------------------------------------------------------+ | | | v v v +----------------------+ +---------------------+ +----------------------+ | 1. Signal Forensics | | 2. Partition Audit | | 3. Label Alignment | +----------------------+ +---------------------+ +----------------------+ | WaveWash: | | SplitSafe: | | DriftCheck: | | Clipping & DC bias | | Acoustic matching | | Timed label drift | | BandCheck: | | Cross-split leaks | | TonicFind: | | Spectral cutoffs | +---------------------+ | Musical key labels | | PhaseCheck: | | TempoCheck: | | Mono downmix loss | | BPM & octave errors | +----------------------+ +----------------------+
Defect Category Acoustic Manifestation Detection Engine Scope Input Requirements
Waveform Damage Digital clipping, rail saturation, DC offset bias wavewash File-level Raw audio (WAV, FLAC)
Spectral Cutoff Brickwall filter drop, fake upsampling bandcheck File-level Raw audio
Phase Cancellation Polarity inversion, mono downmix cancellation phasecheck File-level 2-channel stereo audio
Annotation Drift Time offset, linear drift (ppm), edit jumps driftcheck Pair-level Audio + timed labels (SRT/VTT/CSV)
Split Leakage Transformed duplicates across Train/Val/Test splitsafe Dataset-wide Split folders or manifest
Key Mislabel Mismatched key label or modal confusion tonicfind Pair-level Audio + key label (metadata/filename)
Tempo Mislabel Incorrect BPM, half-time/double-time octave error tempocheck Pair-level Audio + BPM label (metadata/filename)
—————————————————————————————————————————————
- The 7 Diagnostic Tools: Algorithms and Interfaces The ComposerMind AudioIntegrity Suite consists of lightweight Python packages built upon standard numerical foundations (numpy, scipy, soundfile). Each tool addresses a specific stage of the pre-training quality pipeline. 3.1 WaveWash (v0.2.0): Waveform Forensics and Headroom Cleaning • Defect Identified: Digital flat-top clipping, harsh limiter distortion, and non-zero DC displacement. • Mechanism: Scans consecutive sample runs within 0.001 dB of full scale (0 dBFS) and measures mean waveform displacement (E[x(t)] ≠ 0). • Measured Thresholds: Files with ≥ 0.1% clipped samples raise a warning/rejection flag. DC bias is measured as a percentage of full-scale displacement. • Correction Capability: Includes an automated baseline high-pass filter that attenuates sub-audible DC displacement while minimally affecting audible program material. 3.2 BandCheck (v0.1.1): Spectral Cliff Detection and Bandwidth Verification • Defect Identified: Low-bitrate codec compression disguised inside high-sample-rate containers. • Mechanism: Computes a smoothed average power spectrum across the recording using the Short-Time Fourier Transform (STFT). The detector scans from 2,500 Hz upward to identify the steepest spectral drop (a "cliff" of ≥ 25 dB relative to the band's level). • Verdict Rules: • OK (PASS): Cutoff ≥ 90% of Nyquist (or ≥ 19.5 kHz, whichever is lower). • LIMITED (WARN): Cutoff between 60% and 90% of Nyquist (for example, a 16.0 kHz cutoff in a 48 kHz file is 66.7% of the 24 kHz Nyquist, triggering LIMITED). • FAKE (FAIL): Cutoff below 60% of Nyquist (< 14.4 kHz at 48 kHz). • Domain Note: BandCheck is calibrated for full-range broadband audio (speech, ensembles, and modern production). Narrow-band solo instruments (such as an upright bass rolling off sharply at 2.1 kHz) will trigger bandwidth warnings and require domain-informed review. (Note: While BandCheck's search window starts at 2,500 Hz, the reported cutoff frequency reflects the knee where spectral power drops 10 dB below the cliff band, which for the synthetic bass falls at ~2,144 Hz). Frequency Spectrum Visualization (48 kHz / 24 kHz Nyquist): Genuine 48 kHz: [====================================] -> OK (PASS) 16 kHz Cutoff: [========================............] -> LIMITED (WARN) 11 kHz Cutoff: [================....................] -> FAKE (FAIL)
3.3 PhaseCheck (v0.1.0): Spatial Coherence and Downmix Loss
• Defect Identified: Polarity inversions and destructive phase cancellation between stereo channels.
• Mechanism: Measures inter-channel cross-correlation ρ ∈ [-1.0, +1.0] and mono downmix loss in dB over active frames:
Δmono = -10 log₁₀( frac{p{mono}}{(p_L + p_R)/2} ),dB
where p_{mono} is the power of the mono sum M = (L + R) / 2, and (p_L + p_R) / 2 is the mean power of the two channels. Downmix loss is numerically capped at 60.0 dB.
• Verdict Rules:
• PASS: Full-band downmix loss < 6.0 dB. Identical channels (dual mono) lose 0.0 dB and are explicitly noted.
• WARN: Full-band loss ≥ 6.0 dB, or loss in any active sub-band (20–250 Hz, 250–2000 Hz, 2–8 kHz, above 8 kHz) ≥ 10.0 dB.
• FAIL: Full-band loss ≥ 10.0 dB. (Reaching a 10 dB loss requires inverted signals to have had an initial correlation c ≥ 0.8; uncorrelated stereo channels decorrelate rather than completely cancel).
3.4 DriftCheck (v0.1.0): Timed Label Alignment and Offset Analysis
• Defect Identified: Misalignment between timed speech labels (subtitles, transcripts) and corresponding audio files.
• Mechanism: Compares audio voice activity against timestamped intervals across formats including SRT, WebVTT, CSV/TSV, and Lhotse manifests (.jsonl, .jsonl.gz).
• Parameters Measured:
• Constant Offset: Shift in seconds across all annotations.
• Linear Drift Rate: Cumulative timing stretch measured in parts-per-million (ppm) or stretch ratio (15 ppm corresponds to 0.9 ms of drift per minute of audio: 15 times 10^{-6} times 60 s = 0.0009 s).
• Timing Jumps: Discontinuities caused by edits, spliced advertisements, or cut segments.
• Unlabeled Speech / Dead Labels: Speech segments lacking annotations or labels placed over silence.
3.5 SplitSafe (v0.1.1): Acoustic Partition Firewall
• Defect Identified: Cross-partition contamination and transformed duplicates across train, validation, and test splits.
• Mechanism: Extracts constellation peak pairs from audio spectrograms (time-frequency landmarks) and aggregates time-offset alignment votes across split folders.
• Tested Invariance: Invariant to bit-level transformations including exact duplication, temporal trimming, and gain adjustments.
3.6 TonicFind (v0.1.0): Harmonic Key Annotation Verification
• Defect Identified: Mislabeled musical key annotations in music information retrieval corpora.
• Mechanism: Extracts 12-dimensional chromagram features and evaluates cosine correlation against established Krumhansl-Schmuckler and Temperley key profiles. Reports Camelot wheel distance and separates relative major/minor ambiguities from dissonant errors.
3.7 TempoCheck (v0.1.0): Rhythmic and BPM Annotation Verification
• Defect Identified: Mislabeled tempo annotations and downbeat tracking errors.
• Mechanism: Computes spectral flux novelty envelopes and autocorrelation lag peaks. Differentiates octave tempo ambiguities (such as 70 BPM vs. 140 BPM half-time/double-time relationships) from true tempo mismatches.
—————————————————————————————————————————————
- Controlled Benchmark Evaluation (Synthetic Ground Truth) To evaluate detection performance against mathematically known ground truth, a controlled benchmark corpus of 34 audio files at 48,000 Hz (24-bit linear PCM) was synthesized using generate_synthetic_corpus.py (seed: 42) and indexed in ground_truth_manifest.json. The 34-file manifest contains: • 10 Clean Negative Controls: 6 broadband acoustic harmonic signals (clean_00–clean_05) and 4 naturally bandwidth-limited acoustic signals (clean_06–clean_09, synthetic acoustic bass rolling off sharply at 2.1 kHz). • 7 Bandwidth Cutoff Files: Chebyshev Type II brickwall cutoffs at 8 kHz, 11.025 kHz, 12 kHz, 15 kHz, 16 kHz, 18 kHz, and 20 kHz. • 6 Phase Defect Files: 1 full polarity inversion (R = -L) and 5 inter-channel time delays (0.08 ms, 0.25 ms, 0.50 ms, 1.00 ms, 2.00 ms). • 4 Clock Drift Files: Multi-device sample clock drift at 15 ppm, 25 ppm, 50 ppm, and 100 ppm. • 2 Waveform Defect Files: 1 hard digital clipping (≥ 0.1% clipped samples) and 1 non-zero DC bias (+12% full-scale displacement). • 4 Inter-Stem Bleed Files: Multi-track drum-into-vocal bleed (-10 dB to -40 dB) authored for stem separation diagnostics (evaluated in Article 2). • 1 Split Leakage Test Pair: Clean training partition vs. exact and trimmed/gain-modified test copies. (Note: Harmonic key and tempo annotation checkers, TonicFind and TempoCheck, validate metadata tags against audio signals and are evaluated on their respective dedicated test suites of 25 real songs each from their repositories, rather than raw signal forensics files in the audio manifest). Synthetic Benchmark Evaluation Table Diagnostic Tool Target Defect Evaluated Clean Controls Tested Clean False Alarms Defective Cases Tested Detected Cases Precision Recall Measured Benchmark Output wavewash Clipping & +12% DC Offset 10 0 / 10 2 2 100.0% 100.0% DC bias reduced: 12.0% → 0.0% bandcheck Spectral cutoffs (8k–20k Hz) 10 4 / 10 7 6 60.0% 85.7% Cutoff MAE: 37.7 Hz (miss at 20 kHz; 4 FP on bass) phasecheck Polarity inversion & delay 10 0 / 10 6 1 / 6 (3 / 6)* 100.0% 16.7% (50.0%)* Inversion: ρ = -1.00, Δ_mono = 60.0 dB (cap); delays lose < 5 dB driftcheck Clock drift (15–100 ppm) 10 0 / 10 4 4 100.0% 100.0% Drift estimation error < 0.4 ppm splitsafe Train/Test duplicate leakage 10 0 / 10 2 2 100.0% 100.0% Exact (100%) & trimmed/scaled (82% overlap) tonicfind Key annotation mismatch Dedicated suite (25 songs) 0 / 25 25 wrong keys 25 100.0% 100.0% 100% exact match on 25 songs (MIREX 1.000) tempocheck BPM annotation error Dedicated suite (25 songs) 0 / 25 (FAIL) 25 wrong BPM 22 95.7% 88.0% 84.0% Acc1, 100.0% Acc2 (octaves); 4 half/double WARNs
(Note on PhaseCheck detection: 1 of 6 flagged under published package loss rules (≥ 6 dB); 3 of 6 flagged under benchmark runner's extra correlation rule (ρ < 0.2)).
Statistical and Methodological Observations
- Sample Size and Wilson Intervals: For the tools observing 0 false alarms on 10 clean controls (wavewash, phasecheck, driftcheck, splitsafe), the 95% Wilson score confidence interval is [0.0%, 27.8%]. For BandCheck, which produced 4 false alarms on the 4 naturally bandwidth-limited acoustic bass controls, the observed false-alarm rate is 40.0% ([16.8%, 68.7%]). These figures represent baseline sample stability rather than a universal zero false-alarm guarantee.
- Reconciling BandCheck's 60.0% Precision: BandCheck achieved 85.7% recall (detecting 6 of 7 cutoffs, missing only the 20 kHz cutoff which sits near the 90% full-bandwidth ceiling). However, BandCheck flagged the 4 naturally bandwidth-limited acoustic bass controls (clean_06–clean_09) as FAKE (cutoff ~2144 Hz). Because BandCheck is an acoustic-agnostic detector designed for full-spectrum audio, sharp natural instrument roll-offs below 60% of Nyquist (< 14.4 kHz) trigger bandwidth flags, producing 4 false positives (Precision = 6 / (6 + 4) = 60.0%). This highlights the necessity of reviewing bandwidth flags within musical domain context.
- Reconciling PhaseCheck's 1/6 vs. 3/6 Recall: PhaseCheck detected the polarity-inverted file with 60.0 dB loss (maximum cap) and ρ = -1.00. Among the 5 inter-channel time delay files (0.08 ms to 2.00 ms), delay creates comb filtering rather than full-band cancellation; full-band downmix loss remained between 0.5 dB and 4.5 dB, not reaching PhaseCheck's 6.0 dB warning threshold. Under the published package's downmix loss rules (≥ 6 dB), recall is 16.7% (1 of 6). When evaluated with the benchmark runner's secondary correlation check (ρ < 0.2), the two longest delays (1.0 ms and 2.0 ms) were flagged, giving 3 of 6 detected (Recall = 50.0%, Precision = 100.0%). PhaseCheck is designed for destructive phase cancellation, not time-delay estimation.
- SplitSafe Robustness Scope: Within this synthetic benchmark, SplitSafe was evaluated on exact duplication (100% matched landmarks) and temporal trimming with gain scaling (0.1s trim, 0.95 gain multiplier, 82% overlap). It relies on spectral peak pairs; unverified robustness claims regarding severe non-linear pitch-shifting or heavy time-stretching are excluded. —————————————————————————————————————————————
- Real-World Music Validation: MUSDB18-HQ Benchmark To assess phase-cancellation detection in complex multi-track production environments, phasecheck was benchmarked against 200 stereo audio files (stems and mixtures) sampled from the MUSDB18-HQ train split. The results below reflect the pre-registered second evaluation run (eval-spec-pc-v2) conducted on the 200 files of the train split, after an initial exploratory run on the test split revealed that downmix cancellation loss depends fundamentally on initial inter-channel correlation (c). Real-World Music Evaluation Table (MUSDB18-HQ Train Split, 200 Files) Test Condition Input Stems / Files Flagged (WARN or FAIL) Reached FAIL (≥ 10 dB) 95% Wilson Interval Notes on Acoustic Distribution Untouched Studio Originals 200 files 1 flagged (0.5%) 0 (0.0%) [0.1%, 2.8%] WARN on 1 vocals stem (9.7 dB loss, ρ = -0.79; unconfirmed by listening) Planted Polarity Inversion (All Stems) 180 stems 169 flagged (93.9%) 128 (71.1%) [89.4%, 96.6%] Inversion across all stereo stems regardless of initial correlation Inversion Subset: Initial ρ ≥ 0.5 169 stems 168 flagged (99.4%) 128 (75.7%) [96.7%, 99.9%] Post-hoc favorable subset: correlated stereo stems cancel strongly Inversion Subset: Initial ρ < 0.5 11 stems 1 flagged (9.1%) 0 (0.0%) [1.6%, 37.7%] Wide decorrelated stereo loses < 6 dB even when inverted Dual-Mono Recognition (Mono in Stereo) 180 stems 180 identified (100.0%) 0 (0.0% FAIL; 180 PASS) [97.9%, 100.0%] Correctly identified identical channels (0.0 dB loss, marked PASS) 2–8 kHz Band Polarity Inversion 71 stems 62 flagged (87.3%) 0 (0.0%) [77.6%, 93.3%] Sub-band cancellation detected via multi-band analysis (WARN)
Real-Data Insights and Calibration Lessons
• Initial Evaluation Miss: In the exploratory first run (eval-spec-pc-v1) on the test split, the evaluation plan hypothesized that 99% of channel-inverted stems would reach FAIL (≥ 10 dB loss). In reality, only 54% (104/194) reached FAIL. Downmix loss of an inverted stereo pair with initial correlation c follows:
Δ_mono ≈ -10 log₁₀((1 - c) / (2)),dB
Reaching FAIL (≥ 10 dB) requires initial correlation c ≥ 0.8; reaching WARN (≥ 6 dB) requires c ≥ 0.5. Inverted stems with low initial correlation (c < 0.5, 11 stems in the v2 run) lose less than 6 dB and cannot be flagged by downmix loss alone.
• Unconfirmed Original Alert: One natural vocal stem (Giselle - Moss/vocals.wav) triggered a WARN with a 9.7 dB downmix loss and ρ = -0.79. Because listening tests were not performed on this file, it represents an unconfirmed alert rather than a verified false alarm.
—————————————————————————————————————————————
- Reference Corpus Self-Audit: OpenAudio-48k To demonstrate pre-training dataset hygiene in practice, the framework was deployed as an author self-audit during the mastering of the ComposerMind OpenAudio-48k corpus—a 94-track studio-rendered electronic and interactive audio research dataset hosted on Zenodo CERN: • Waveform Verification (
wavewash): All 94 tracks were audited for clipping and DC offset. Audits confirmed 0 of 94 tracks with clipped samples and 0.0% DC displacement. • Bandwidth Audit (bandcheck): All 94 tracks were verified to contain full-band spectral energy extending to the 24 kHz Nyquist ceiling, with 94 of 94 tracks showing no spectral cutoff below 19.5 kHz. • Spatial Compatibility (phasecheck): Audited all stereo stems for mono downmix compatibility; 0 of 94 tracks were flagged for downmix cancellation (Δ_mono < 6.0 dB). ————————————————————————————————————————————— - Installation, CLI Usage, and Pipeline Integration 7.1 PyPI Package Installation Each diagnostic utility is independently available on PyPI with pinned releases: pip install wavewash==0.2.0 \ splitsafe==0.1.1 \ bandcheck==0.1.1 \ phasecheck==0.1.0 \ driftcheck-audio==0.1.0 \ tempocheck==0.1.0 \ tonicfind==0.1.0
7.2 Command-Line Execution
Run forensic audits across unverified directories:
1. Audit folder for fake 48 kHz upsampling and codec cutoffs
bandcheck check ./scraped_corpus/ --out ./reports/bandwidth_audit.csv
2. Check for train/validation data leakage across partitions
splitsafe check ./data/train/ ./data/val/ --out ./reports/splitsafe_report/
3. Clean DC offset and verify headroom
wavewash run --input ./raw_recordings/ --output ./sanitized_recordings/
4. Check stereo files for destructive mono downmix cancellation
phasecheck check ./stereo_stems/ --out ./reports/phase_audit.csv
5. Check timed speech labels against audio alignment
driftcheck check ./speech_corpus/ --out ./reports/drift_report/
7.3 Programmatic Ingestion Gate (Python Data Loader)
Integrate quality gates directly into ingestion or data preparation pipelines:
from pathlib import Path
from bandcheck.check import check_file as check_bandwidth
from phasecheck.check import check_file as check_phase
def audit_audio_sample(audio_path: Path) -> bool:
"""Pre-training quarantine and review filter (flags suspicious files for inspection)."""
# Step 1: Flag lossy-compressed or fake-upsampled audio
spectral_result = check_bandwidth(audio_path)
if spectral_result.status == "FAKE":
# Cutoff is below 60% of Nyquist; flag file for human review (not automatic deletion)
return False
# Step 2: Flag stereo files with catastrophic mono downmix cancellation phase_result = check_phase(audio_path) if phase_result.status == "FAKE" or ( phase_result.loss_db is not None and phase_result.loss_db >= 10.0 ): # Downmix loss exceeds 10 dB threshold; flag file for human review return False return True —————————————————————————————————————————————
Practical Action Guide: Handling Flagged Files
Automated audits should not automatically discard or modify data. Instead, flagged files should be triaged through human-in-the-loop review:
+-------------------------------------------------------------------------+
| FLAGGED FILE TRIAGE |
+-------------------------------------------------------------------------+
| | |
v v v
[ DC Bias / Headroom ] [ Codec Cutoff / Phase ] [ Leakage / Drift ]
| | |
v v v
Review & filter on copy Listen to mono mix first Re-partition splits
(Neutralize DC offset) (Quarantine if confirmed) Correct label offsetRepairable Signal Defects (DC Offset, Gain Headroom): High-pass filtering attenuates sub-audible DC displacement while minimally affecting audible program material. Process through wavewash run on a duplicate copy to neutralize DC bias.
Phase Inversions: Always listen to the mono downmix first. If a stereo file flags FAIL (≥ 10 dB loss) and critical listening confirms that the acoustic signal collapses or sounds hollow in mono, invert one channel on a copy (R' = -R) and re-audit. Do not invert channels without listening verification, as naturally decorrelated or antiphase stereo material (such as wide ambient recordings or the Giselle vocal stem) would be degraded by polarity inversion.
Spectral Cutoffs (
LIMITED/FAKE): Inspect the recording domain. If the file is a solo acoustic instrument (such as upright bass), retain it. If it is full-spectrum music or speech with an artificial brickwall cutoff below the model's target bandwidth, quarantine it from the training corpus.Partition Leakage: Move identified duplicate pairs into the same split partition (e.g., place all variations in Train, removing them from Validation and Test).
—————————————————————————————————————————————Framework Limitations and Failure Cases
A dependable quality assurance tool must make its scientific boundaries clear:Labelled Real-Data Benchmark Scope: In this article, only phasecheck is evaluated on an independent, labelled real-world multi-track music benchmark (MUSDB18-HQ). Section 6 reports an internal self-audit of our own OpenAudio-48k corpus (confirming zero flags on clean master exports, which serves as author verification rather than independent labelled validation). The remaining diagnostic tools (splitsafe, driftcheck, tonicfind, tempocheck) are evaluated on controlled synthetic benchmarks, with their independent real-data checks documented in their dedicated sub-repositories.
Channel Swaps and Time Delays (
phasecheck): PhaseCheck measures mono downmix cancellation loss. It cannot detect swapped left/right channels, nor can it reliably detect inter-channel time delays (small delays lose < 5 dB; while the benchmark runner's correlation check flagged 1 ms and 2 ms delays, PhaseCheck contains no dedicated delay estimator).Narrow-Band Acoustic Audio (
bandcheck): BandCheck detects spectral drop-offs relative to container sample rate. Solo recordings of bass, speech, or naturally muted instruments lack high-frequency content and will trigger bandwidth flags (LIMITED or FAKE) despite being pristine uncompressed recordings.Uncorrelated Stereo Channels (
phasecheck): Polarity inversion detection requires correlated channels (c ≥ 0.5). In wide, decorrelated stereo material (such as ambient stereo field recordings), polarity inversion produces minimal cancellation loss (< 3 dB) and will not be caught.SplitSafe Scope: Within the synthetic evaluation, SplitSafe was tested on exact duplicates (100% match) and a 0.1s trim with 0.95 gain scaling (82% overlap). It relies on spectral peak pairs; unverified robustness claims regarding severe non-linear pitch-shifting or heavy time-stretching are excluded.
Annotation Dependence (
driftcheck,tonicfind,tempocheck): These tools validate existing metadata against audio waveforms. They cannot infer missing ground-truth annotations from scratch.
—————————————————————————————————————————————Reproducibility Trail
To enable independent verification, all benchmark specifications, synthetic generators, and evaluation logs are publicly archived:
• Synthetic Benchmark Generator: Contained in generators/generate_synthetic_corpus.py (seed: 42, 48 kHz, 24-bit linear PCM).
• Ground Truth Manifest: Documented in results/ground_truth_manifest.json with relative paths (corpus/...) across all 34 synthetic corpus items.
• Real-Data Verification Scripts: Evaluation scripts are available in their respective repositories: PhaseCheck/scripts/real_data_check.py, TonicFind/scripts/real_data_check.py, and Tempocheck/scripts/real_data_check.py.
• Suite Architecture: The 7 dataset QA tools form part of the broader 10-tool ComposerMind research initiative (with the remaining 3 post-training evaluation tools documented in Article 2).
• Repository Releases & Tags: Code and evaluation plans archived at GitHub release tags v0.1.0 and v0.1.1 in ComposerMind-AudioIntegrity-Suite, with PhaseCheck evaluation plans tagged as eval-spec-pc-v1 and eval-spec-pc-v2 in the PhaseCheck repository.
—————————————————————————————————————————————Conclusion
Clean training data is the foundation of dependable audio AI models. By replacing ad-hoc spot-checks with systematic, open-source signal forensics, partition isolation, and annotation verification, the ComposerMind AudioIntegrity Suite enables machine learning practitioners to identify data defects before allocating GPU compute.
All seven tools are open-source, available on PyPI, and designed for modular integration into existing data preparation workflows.
—————————————————————————————————————————————
Resources and Repository Links
• PyPI Distribution: pip install wavewash splitsafe bandcheck phasecheck driftcheck-audio tempocheck tonicfind
• GitHub Monorepo: github.com/BoobalanArjunan/ComposerMind-AudioIntegrity-Suite
• Research Archive: ComposerMind OpenAudio-48k (Zenodo CERN)
• Author Affiliation: Boobalan Arjunan, Lead Audio AI Architect & Researcher, ComposerMind Labs, Chennai, India.
