Total Audio Files
40
10 voices × 4 sets
Best SNR (Cleaned)
76.6 dB
DigiTek cleaned avg
SNR Improvement
+10.2 dB
DigiTek set avg gain
Best P808 MOS
3.74/5
DigiTek cleaned
Clipped Files
10
DigiTek originals only
Avg Clipped Samples
58.9
DigiTek orig per file
Summary Comparison — All 4 Sets
| Metric | BoAt Original | BoAt Cleaned | DigiTek Original | DigiTek Cleaned |
|---|---|---|---|---|
| SNR (dB) | 41.6 | 71.7 ↑+30.1 | 44.3 | 76.6 ↑+32.3 |
| RMS (dBFS) | -16.7 | -18.0 | -13.3 | -19.8 |
| Clipping % | 0% | 0% | 0.031% | 0% |
| OVRL MOS | 2.963 | 3.088 | 2.981 | 3.257 |
| SIG MOS | 3.398 | 3.397 | 3.348 | 3.509 |
| BAK MOS | 3.705 | 3.980 | 3.765 | 4.095 |
| P808 MOS | 3.471 | 3.637 | 3.613 | 3.738 |
| Sample Rate | 44,100 Hz · Mono · 16-bit PCM | |||
📌 BoAt Mic — Key Findings
No clipping on any of the 10 files. Gain was set appropriately with sufficient headroom.
Background noise floor ~-46 dBFS. High-frequency hiss (~8 kHz). SNR averaged 29.5 dB before cleaning (below 40 dB TTS threshold).
⚠️ DigiTek Recorder — Key Findings
All 10 files clipped. voice_1 worst with 123 clipped samples. Gain was set too high on the budget recorder (no limiter).
Despite clipping, DigiTek cleaned achieves best scores overall. Fix: lower gain 6–10 dB for future sessions.
🎯 DNSMOS Threshold
Minimum acceptable for TTS: 3.5 on all scores. Only DigiTek cleaned meets SIG MOS ≥3.5. BAK MOS improved most from cleaning (+0.27–0.33 across sets).
✅ Cleaning Result
Noise reduced 85% (prop_decrease=0.85). SNR improved +8.6 dB (BoAt) and +10.2 dB (DigiTek). Peak normalised to -3 dBFS. Originals untouched.
Audio File Browser & Player
SNR Analysis — Before vs After Cleaning
SNR Before vs After Cleaning
BoAt mic — all 10 files
RMS Level Before vs After
BoAt mic — dBFS values
Per-File Metrics — MotoMimi_BoAt
| File | SNR Before | SNR After | Improvement | RMS Before | RMS After | Clipped |
|---|
DNSMOS P.835 — Perceptual Quality Scores
About DNSMOS: Deep learning model by Microsoft (DNS Challenge) that predicts perceptual audio quality without a clean reference.
Trained to approximate human listener ratings. All scores on a 1–5 scale. Minimum acceptable for TTS: ≥3.5.
Input audio resampled to 16 kHz mono.
DNSMOS — All Scores Comparison
Average scores across all 4 sets
Improvement from Cleaning
Delta scores (cleaned − original)
📖 Score Definitions
- OVRL MOS — Combined perceptual quality (signal + background)
- SIG MOS — Clarity and naturalness of speech
- BAK MOS — Background quality (higher = cleaner noise)
- P808 MOS — ITU-T P.808 standard overall estimate
💡 Key Observations
- BAK MOS improved most after cleaning (+0.27–0.33)
- SIG MOS is similar across sets (~3.35–3.51) — speaker quality is consistent
- DigiTek cleaned achieves the best scores overall
- OVRL MOS modest (2.96–3.26) — real-world recording conditions
File Explorer
Cleaning Process — Technical Details
🔬 Step 1 — Noise Profile Estimation
Each file was divided into 0.3s non-overlapping windows. The window with lowest RMS energy was identified as the noise profile. This per-file approach captures the unique noise character of each recording.
🌊 Step 2 — Spectral Noise Reduction
Library: noisereduce (Python v2.x) · Non-stationary spectral gating
- STFT of signal and noise profile
- Frequency-dependent threshold mask
- prop_decrease = 0.85 (avoids musical artefacts)
- stationary = False (adapts to noise drift)
📏 Step 3 — Peak Normalisation
Each cleaned file normalised to -3 dBFS peak amplitude (~70.8% of max). Ensures consistent loudness, prevents new clipping, leaves 3 dB headroom.
y = y × (10^(−3/20) / max(|y|))
💾 Step 4 — Export
Files saved as 16-bit PCM WAV at original sample rate (44,100 Hz). Names simplified to voice_N_cleaned.wav. Originals never modified.
Recording Setup
| Parameter | MotoMimi_BoAt | MotoMimi_Studio_DigiTek |
|---|---|---|
| Microphone | BoAt mic | DigiTek recorder (built-in) |
| Sample Rate | 44,100 Hz | 44,100 Hz |
| Channels | Mono | Mono |
| Bit Depth | 16-bit PCM | 16-bit PCM |
| Files | 10 (voice_1–10) | 10 (voice_1–10) |
| Clipping | None | Yes (all 10) |
| Noise Floor | ~-46 dBFS | ~-45 dBFS |
🔧 Fix for DigiTek Going Forward
Lower recording gain by 6–10 dB so peaks stay below -6 dBFS. This gives headroom for plosives and loud moments. Target RMS: -18 to -22 dBFS.
The DigiTek has no limiter or auto-gain control. A professional recorder (e.g. Zoom H5) would protect against peaks. Until upgrade, manual gain control is essential.