sw-voice/swamiji-tts-cfgskip-ab

CFG-skip A/B — judge the failed configs by ear

⚠️ SYNTHETIC AUDIO

Every clip is generated by a text-to-speech model. Nothing here is a recording of any person. The passages are neutral benchmark text — not transcripts, quotations, or teachings of Sri Ganapathi Sachidananda Swamiji or anyone else.

Four configs, the same 20 passages with the same per-request seeds, so the only variable is the guidance strategy. a_baseline/3.wav and b_cfg_k3/3.wav are the same sentence — play them back to back.

These are the configs that reached n=75–80 concurrent but which I flagged as failing on quality. Published so the call is yours rather than mine.

What you're comparing

folder	config	max concurrency @ RTF ≤ 0.875
a_baseline/	full CFG on all 9 Euler steps — the reference	64
b_cfg_k3/	full CFG on first 3 steps, cached delta for 6	75–80
c_cfg_k1/	full CFG on step 1 only, cached delta for 8	80
d_cfg_k3_scaled/	as b, delta rescaled by field decay	75–80

Why I called it a fail

Half the diffusion head's batch is the unconditional pass that classifier-free guidance needs. Skipping it on later steps and reusing the cached guidance delta halves those steps' cost — that's where the extra concurrency comes from.

The numbers say intelligibility is fine and level uniformity is better than baseline:

	dBFS	level sd	median WER	mean WER	crest factor
a_baseline	−17.77	4.26	0.014	0.0141	16.3 dB
b_cfg_k3	−11.66	2.33	0.014	0.0169	11.4 dB
c_cfg_k1	−12.24	2.65	0.014	0.0133	11.9 dB
d_cfg_k3_scaled	−10.35	1.28	0.013	0.0134	10.2 dB

(Crest = mean peak / mean RMS across the 20 clips in each set, recomputed directly from these files.)

WER is unchanged or better. Level consistency is up to 3.3× tighter. So by two of the three metrics this is an improvement.

The problem is the crest factor: 16.3 dB → 10.2 dB. Nothing clips (peaks sit at 0.996, zero clipped samples), but every clip pins its peak at full scale while RMS rises ~7 dB. That is dynamic-range compression, not a gain offset — turning the volume down would match the loudness and leave the dynamics squashed. Speech with 7 dB less dynamic range tends to read as flat, loud, and close-miked.

That is precisely the kind of defect WER cannot see, which is why the level and crest metrics exist here at all: an earlier attempt (inference_timesteps=4) passed a WER check while being 7 dB quiet and audibly degraded.

What to listen for: not words — those are intact. Listen for dynamics. Does the baseline breathe between phrases where the CFG-skip versions feel pushed forward and even? If that difference doesn't bother you at your playback levels, n=75–80 is available and I'll wire it up.

d_cfg_k3_scaled was my attempted fix — rescaling the cached delta by how much the velocity field had decayed. It made the compression worse (−10.35 vs −11.66 dBFS), which is how I learned the field grows rather than shrinks across Euler steps.

Config common to all four

model:               sw-voice/swamiji-voxcpm2-ft-lr0.0001-10ep-deva
engine:              nano-vllm-voxcpm 2.0.3 + PR#79 (streaming VAE)
inference_timesteps: 10          # unchanged in every variant
cfg_value:           1.5
max_model_len:       400
compile_dit:         true
fp16_dit:            inductor    # fp16-accumulate GEMMs, separately verified clean
GPU:                 1x RTX 3080 12 GB

Only cfg_full_steps differs: 9 (baseline), 3, 1, 3-scaled.

Contents

Each folder holds 0.wav … 19.wav (48 kHz mono, 16-bit) plus meta.json with the per-clip text, duration and dBFS. Same index = same passage and seed across all four folders.

For reference: the config that passed

sw-voice/swamiji-tts-64-concurrent-fp16 — 64 concurrent, median WER 0.014, level within 0.14 dB of baseline, crest factor unchanged. That is the one I'd ship without asking.