sotto

sotto encoder v0: first oracle results

Written 2026-10-31. Code: sim/oracle_ladder.py, sim/vocoder_lab.py, sim/encoder_race.py. Results: docs/results/oracle_ladder_15s.json and docs/results/encoder_race.json. Samples: docs/samples/encoder/.

Update: v1 (octave-corrected pitch) is in encoder-v1.md; the listening page now uses v1.

Where Codec 2 2400 loses quality

c2sim was run on the three 15 s clips with one simplification added at each rung:

Rung	Female	Male (LibriSpeech)	Male (ve9qrp)
Sinusoidal model, true phases	3.08	3.82	4.00
+ synthetic phase (voicing)	2.78	3.52	3.27
+ postfilter	2.77	3.54	3.48
+ LPC-10 envelope	2.61	2.98	2.94
+ scalar LSP quantization	2.50	2.87	2.80
+ 2400 pitch/energy quantization, 20 ms	2.33	2.90	2.63
c2enc/c2dec 2400	2.16	2.91	2.85

For the men, the 10th-order LPC envelope is the largest single loss, about 0.55 PESQ. Synthetic phase costs about 0.3 to 0.7.
The female clip starts low: 3.08 even with true phases. That points at the analysis itself (pitch tracking, or too few harmonics to sample her envelope) more than at quantization.
Higher LPC orders barely help (16: +0.1–0.2; 20: about the same). A mel-spaced “rate-K” envelope (20 amplitudes from 200 to 3700 Hz, as in Codec 2’s newamp1) recovers most of the loss when unquantized: 2.67 / 3.43 / 3.31 against 2.77 / 3.54 / 3.48 with no envelope modelling at all.

sotto v0

v0 is a 48-bit, 20 ms frame with no memory between frames:

Field	Bits	Notes
Voicing	2	one decision per 10 ms, from Codec 2’s estimator
Pitch (Wo)	7	log-uniform, Codec 2’s range
Frame energy	5	sum of harmonic powers, 10–80 dB; the decoded level matches the input
Envelope shape	34	DCT of the mean-removed 20-point mel envelope. Uniform scalar quantizers, with step sizes and a greedy bit allocation (5,4,3,3,2,…) trained on Codec 2’s test recordings. The ve9qrp test window is excluded from training, and the LibriSpeech voices are never used.

The decoder rebuilds harmonic amplitudes from the envelope and applies a newamp1-style formant postfilter (gain 1.5). It interpolates the 10 ms subframe and then uses Codec 2’s synthesis (c2sim --modelin --phase0 --postfilter).

Inter-frame prediction of the shape (α = 0.5) gave a further +0.1 to 0.3 on two clips. It was left out of v0 because channel errors would carry into the next frame.

Encoder against encoder

Both codecs send 48 bits every 20 ms through the identical chain: sotto VW payload, C4FM softmodem, the same fading seeds, and the same receiver with CRC concealment. PESQ is averaged over 3 seeds:

Condition	Codec	Female	Male (LibriSpeech)	Male (ve9qrp)
No radio channel	Codec 2 2400	2.16	2.91	2.85
	sotto v0	2.50	3.14	2.88
446 MHz 60 km/h, 52 dB-Hz	Codec 2 2400	1.69	2.29	2.49
	sotto v0	2.11	2.68	2.69
446 MHz 60 km/h, 56 dB-Hz	Codec 2 2400	2.16	2.72	2.78
	sotto v0	2.45	3.05	2.85
146 MHz 30 km/h, 56 dB-Hz	Codec 2 2400	1.82	2.06	1.91
	sotto v0	2.19	2.61	2.31

sotto v0 is ahead in every condition and voice: +0.03 to +0.34 with no channel, and +0.2 to +0.55 over fading channels. The fading gains may partly come from v0 being memoryless and its frames concealing more gracefully. The female voice gains the most with no channel (+0.34). She is still well below the men, which leaves pitch and voicing as the next targets.

Caveats

PESQ only, on three clips, and no listening panel yet. The formant postfilter may flatter PESQ. The level difference between the two codecs is under 1 dB.
One training speaker (ve9qrp) is also a test speaker, though a different stretch of the recording. The LibriSpeech voices are fully held out.
v0 is a Python and c2sim prototype, not yet a C encoder and decoder. Its analysis (pitch, voicing) and synthesis are still Codec 2’s.

Pitch tracking and voicing for higher voices: check Codec 2’s estimator on the female clip, and try mixed or band voicing (MBE/MELP-style).
Better shape quantization: trained VQ or predictive coding with error-resilient resets, which needs more training speech.
Port v0 to C inside codec/codec2/ as a new mode, so it runs in real time and on the DSP.

Correction (2026-10-31): ESTOI here was computed with a per-clip waveform alignment that misjudged the codec delay, so it is understated, and PESQ is off by up to ±0.1. Rankings hold. Corrected numbers are in encoder-v1.md.

This site is open source. Improve this page.