Audio Voice Recovery

Forensic Audio Research Audio Voice Recovery Best Practices

Comprehensive audio forensics and voice recovery guide providing CSI-level capabilities for recovering voice from low-quality, low-volume, or damaged audio recordings. Contains 45 rules across 8 categories, prioritized by impact to guide audio enhancement, forensic analysis, and transcription workflows.

When to Apply

Reference these guidelines when:

• Recovering voice from noisy or low-quality recordings
• Enhancing audio for transcription or legal evidence
• Performing forensic audio authentication
• Analyzing recordings for tampering or splices
• Building automated audio processing pipelines
• Transcribing difficult or degraded speech

Rule Categories by Priority

Quick Reference

1. Signal Preservation & Analysis (CRITICAL)

• signal-preserve-original - Never modify original recording
• signal-lossless-format - Use lossless formats for processing
• signal-sample-rate - Preserve native sample rate
• signal-bit-depth - Use maximum bit depth for processing
• signal-analyze-first - Analyze before processing

2. Noise Profiling & Estimation (CRITICAL)

• noise-profile-silence - Extract noise profile from silent segments
• noise-identify-type - Identify noise type before reduction
• noise-adaptive-estimation - Use adaptive estimation for non-stationary noise
• noise-snr-assessment - Measure SNR before and after
• noise-avoid-overprocessing - Avoid over-processing and musical artifacts

3. Spectral Processing (HIGH)

• spectral-subtraction - Apply spectral subtraction for stationary noise
• spectral-wiener-filter - Use Wiener filter for optimal noise estimation
• spectral-notch-filter - Apply notch filters for tonal interference
• spectral-band-limiting - Apply frequency band limiting for speech
• spectral-equalization - Use forensic equalization to restore intelligibility
• spectral-declip - Repair clipped audio before other processing

4. Voice Isolation & Enhancement (HIGH)

• voice-rnnoise - Use RNNoise for real-time ML denoising
• voice-dialogue-isolate - Use source separation for complex backgrounds
• voice-formant-preserve - Preserve formants during pitch manipulation
• voice-dereverb - Apply dereverberation for room echo
• voice-enhance-speech - Use AI speech enhancement services for quick results
• voice-vad-segment - Use VAD for targeted processing
• voice-frequency-boost - Boost frequency regions for specific phonemes

5. Temporal Processing (MEDIUM-HIGH)

• temporal-dynamic-range - Use dynamic range compression for level consistency
• temporal-noise-gate - Apply noise gate to silence non-speech segments
• temporal-time-stretch - Use time stretching for intelligibility
• temporal-transient-repair - Repair transient damage (clicks, pops, dropouts)
• temporal-silence-trim - Trim silence and normalize before export

6. Transcription & Recognition (MEDIUM)

• transcribe-whisper - Use Whisper for noise-robust transcription
• transcribe-multipass - Use multi-pass transcription for difficult audio
• transcribe-segment - Segment audio for targeted transcription
• transcribe-confidence - Track confidence scores for uncertain words
• transcribe-hallucination - Detect and filter ASR hallucinations

7. Forensic Authentication (MEDIUM)

• forensic-enf-analysis - Use ENF analysis for timestamp verification
• forensic-metadata - Extract and verify audio metadata
• forensic-tampering - Detect audio tampering and splices
• forensic-chain-custody - Document chain of custody for evidence
• forensic-speaker-id - Extract speaker characteristics for identification

8. Tool Integration & Automation (LOW-MEDIUM)

• tool-ffmpeg-essentials - Master essential FFmpeg audio commands
• tool-sox-commands - Use SoX for advanced audio manipulation
• tool-python-pipeline - Build Python audio processing pipelines
• tool-audacity-workflow - Use Audacity for visual analysis and manual editing
• tool-install-guide - Install audio forensic toolchain
• tool-batch-automation - Automate batch processing workflows
• tool-quality-assessment - Measure audio quality metrics

Essential Tools

Workflow Scripts (Recommended)

Use the bundled scripts to generate objective baselines, create a workflow plan, and verify results.

• scripts/preflight_audio.py - Generate a forensic preflight report (JSON or Markdown).
• scripts/plan_from_preflight.py - Create a workflow plan template from the preflight report.
• scripts/compare_audio.py - Compare objective metrics between baseline and processed audio.

Example usage:

# 1) Analyze and capture baseline metrics
python3 skills/.experimental/audio-voice-recovery/scripts/preflight_audio.py evidence.wav --out preflight.json

# 2) Generate a workflow plan template
python3 skills/.experimental/audio-voice-recovery/scripts/plan_from_preflight.py --preflight preflight.json --out plan.md

# 3) Compare baseline vs processed metrics
python3 skills/.experimental/audio-voice-recovery/scripts/compare_audio.py \
  --before evidence.wav \
  --after enhanced.wav \
  --format md \
  --out comparison.md

Forensic Preflight Workflow (Do This Before Any Changes)

Align preflight with SWGDE Best Practices for the Enhancement of Digital Audio (20-a-001) and SWGDE Best Practices for Forensic Audio (08-a-001). Establish an objective baseline state and plan the workflow so processing does not introduce clipping, artifacts, or false "done" confidence. Use scripts/preflight_audio.py to capture baseline metrics and preserve the report with the case file.

Capture and record before processing:

• Record evidence identity and integrity: path, filename, file size, SHA-256 checksum, source, format/container, codec
• Record signal integrity: sample rate, bit depth, channels, duration
• Measure baseline loudness and levels: LUFS/LKFS, true peak, peak, RMS, dynamic range, DC offset
• Detect clipping and document clipped-sample percentage, peak headroom, exact time ranges
• Identify noise profile: stationary vs non-stationary, dominant noise bands, SNR estimate
• Locate the region of interest (ROI) and document time ranges and changes over time
• Inspect spectral content and estimate speech-band energy and intelligibility risk
• Scan for temporal defects: dropouts, discontinuities, splices, drift
• Evaluate channel correlation and phase anomalies (if stereo)
• Extract and preserve metadata: timestamps, device/model tags, embedded notes

Procedure:

1. Prepare a forensic working copy, verify hashes, and preserve the original untouched.
2. Locate ROI and target signal; document exact time ranges and changes across the recording.
3. Assess challenges to intelligibility and signal quality; map challenges to mitigation strategies.
4. Identify required processing and plan a workflow order that avoids unwanted artifacts. Generate a plan draft with scripts/plan_from_preflight.py and complete it with case-specific decisions.
5. Measure baseline loudness and true peak per ITU-R BS.1770 / EBU R 128 and record peak/RMS/DC offset.
6. Detect clipping and dropouts; if clipping is present, declip first or pause and document limitations.
7. Inspect spectral content and noise type; collect representative noise profile segments and estimate SNR.
8. If stereo, evaluate channel correlation and phase; document anomalies.
9. Create a baseline listening log (multiple devices) and define success criteria for intelligibility and listenability.

Failure-pattern guardrails:

• Do not process until every preflight field is captured.
• Document every process, setting, software version, and time segment to enable repeatability.
• Compare each processed output to the unprocessed input and assess progress toward intelligibility and listenability.
• Avoid over-processing; review removed signal (filter residue) to avoid removing target signal components.
• Keep intermediate files uncompressed and preserve sample rate/bit depth when moving between tools.
• Perform a final review against the original; if unsatisfactory, revise or stop and report limitations.
• If the request is not achievable, communicate limitations and do not declare completion.
• Require objective metrics and A/B listening before declaring completion.
• Do not rely solely on objective metrics; corroborate with critical listening.
• Take listening breaks to avoid ear fatigue during extended reviews.

Quick Enhancement Pipeline

# 1. Analyze original (run preflight and capture baseline metrics)
python3 skills/.experimental/audio-voice-recovery/scripts/preflight_audio.py evidence.wav --out preflight.json

# 2. Create working copy with checksum
cp evidence.wav working.wav
sha256sum evidence.wav > evidence.sha256

# 3. Apply enhancement
ffmpeg -i working.wav -af "\
  highpass=f=80,\
  adeclick=w=55:o=75,\
  afftdn=nr=12:nf=-30:nt=w,\
  equalizer=f=2500:t=q:w=1:g=3,\
  loudnorm=I=-16:TP=-1.5:LRA=11\
" enhanced.wav

# 4. Transcribe
whisper enhanced.wav --model large-v3 --language en

# 5. Verify original unchanged
sha256sum -c evidence.sha256

# 6. Verify improvement (objective comparison + A/B listening)
python3 skills/.experimental/audio-voice-recovery/scripts/compare_audio.py \
  --before evidence.wav \
  --after enhanced.wav \
  --format md \
  --out comparison.md

How to Use

Read individual reference files for detailed explanations and code examples:

• Section definitions - Category structure and impact levels
• Rule template - Template for adding new rules

Reference Files