The Audio quality controls apply to the next generated take. Existing files stay unchanged. Defaults are 16-bit WAV, broadcast mastering and each engine's best-quality voice refinement (see the table below); model-specific limits still apply.
The compact slider sits below the script in Clone and Design. More options reveals voice refinement (sampling steps), volume balancing and format guidance. The three plain-language choices are Standard, For editing, and Original; the collapsed details explain the WAV formats. Size estimates use decimal MB per minute from the active engine’s output metadata. For unknown model formats, the UI shows the actual size after generation instead of guessing. Voice controls exposes speed, duration and cleanup; technical model controls are collapsed under Advanced model tuning with plain-language names and guidance.
Both popover headers have a Reset to defaults icon. Audio quality resets WAV precision, sampling steps and mastering; Voice controls resets speed, duration, cleanup and advanced model tuning. Each reset preserves the other panel's settings, script, language and selected voice. Audio-quality reset is disabled while generation is running. Existing audio files are never modified.
| WAV slider | Use | Approximate size per minute, 24 kHz mono |
|---|---|---|
| Standard (16-bit PCM) | Normal playback, smallest WAV, broad player support | 2.9 MB |
| For editing (24-bit PCM) | More precision for editing | 4.3 MB |
| Original (32-bit float) | Highest supported export precision | 5.8 MB |
These are uncompressed WAVs. Size scales with duration, sample rate and channel count; 48 kHz doubles these estimates. The finished take shows its actual size, sample rate, channels and sample format. Export precision cannot add information that an engine has already discarded or improve voice identity by itself. The WAV slider does not change model weights, quantization or sample rate.
Voice refinement controls model computation separately. More passes sound cleaner but take longer, without increasing WAV size. Until you move the slider it follows the active engine's best-quality value, shown with a Best quality badge. Below that value the composer pill shows Faster, and Restore best quality returns to it. The slider appears only for adapters that forward this setting:
| Engine | Sampling slider | Best quality (default) | Export controls |
|---|---|---|---|
| OmniVoice, including subprocess mode | 8–64 steps | 32 | 16/24/32-bit WAV |
| VoxCPM2 | 8–64 steps | 32 | 16/24/32-bit WAV, native 48 kHz |
| dots.tts | 8–64 steps | 32 | 16/24/32-bit WAV |
| Supertonic-3 | 5–12 steps, matching its adapter limits | 12 | 16/24/32-bit WAV |
| Other TTS engines | Engine manages sampling; no ineffective slider | — | 16/24/32-bit WAV |
Settings saved by an earlier version that still hold the old fast default (16) move to best quality once, on the first start of this version on that computer; any other value you chose is kept, and values Cloud sync brings in later are never migrated.
After each take the app remembers, on this computer only, the script length, audio length, render time and refinement used, per engine and device. Nothing is sent anywhere. From those measurements it estimates the next take. When the estimate is at least 45 seconds, or at least 15 seconds and three times longer than the audio, a note by Generate shows the expected time and a faster option, for example "About 2 min at best quality · Faster: about 40 s · Refinement 32 → 16". The same note follows a take that was that slow.
The faster option halves refinement (32 → 16 → 8; Supertonic-3 12 → 6 → 5) and picks the mildest step that ends the slow wait, using the measured scaling of your own renders. It is offered only when it saves at least a quarter of the time and five seconds. Volume balancing, cleanup and export format are never lowered. Settings change only when you click Use faster settings; Keep best quality hides the note for this session and Don't ask again turns it off for that engine. Takes that waited for a model to load are not measured. Resetting UI preferences (Reset & remove) clears the measurements and the don't-ask choices.
Even out volume adds the app's processing chain: the gentle mastering pre-stage, then a slow leveler that brings quiet and loud passages closer together, a light EQ and a look-ahead peak limiter that only touches samples above -1 dBFS. The leveler follows whole phrases, not syllables, so peaks keep their shape and pauses and room tails are not pulled up toward speech level. Turning it off skips the app's added EQ, leveling and normalization; engine postprocessing and the existing provenance watermark setting remain in effect. An engine that performs its own mastering continues to skip the app's mastering pre-stage.
The final playback response uses the saved WAV itself. Streaming previews still use PCM16; their completed take uses the selected precision. Engines that only deliver PCM16 cannot recover extra detail through a larger export. OmniVoice and VoxCPM2 subprocesses negotiate float32 transport; legacy PCM16 responses remain supported without reinstalling engines. Float-capable responses have a bounded 128 MiB frame allowance, preserving the old PCM16 duration limit; other engines and requests retain their 64 MiB limit. Retention protects the take being generated even when starred history entries already fill the cap. Cancelled or failed save/read operations remove unpublished WAVs after the writer finishes.
Use Check audio below a finished take to scan locally for long silence, relative volume changes, clipping, empty audio or invalid samples. Click a warning to seek to its timestamp; close the report to dismiss it. Scanning is bounded to two hours and 100 warnings, and reports when limited. It never modifies audio. Warnings are advisory signal checks, not a voice-similarity or naturalness score.
PR #2406 exercises real installed engines on Windows with an RTX 4090:
| Engine | Real generation coverage | Format |
|---|---|---|
| OmniVoice | TTS and cloning, 16/32/64 steps | 24 kHz mono |
| OmniVoice subprocess | TTS and cloning, 16/32/64 steps | 24 kHz mono |
| VoxCPM2 subprocess | TTS and cloning, 16/32/64 steps | 48 kHz mono |
| KittenTTS | Fixed-voice TTS; cloning/step tuning unsupported | 24 kHz mono |
Each condition is exported at all three precisions, with broadcast and raw processing. The first sweep produced 114 valid WAVs with finite samples and no signal warnings. Offline Whisper large-v3 recovered all 13 words in each of the 19 raw float conditions. That short single-text check establishes intelligibility for these trials, not a perceptual ranking across voices or languages.
The sweep exposed seed propagation gaps in the explicit OmniVoice and VoxCPM2 sidecars. Those now receive the take's seed with every chunk, with a fresh sweep after the fix. On the seeded VoxCPM2 clone trial, 16/32/64 steps took approximately 9.7/14.1/27.2 seconds; increasing sampling effort has a measurable time cost.
Model-free regression tests cover export precision, identical final playback/save bytes, streamed final samples, stereo/rate preservation, float sidecar round trips, legacy frames, seed forwarding, warning bounds, path confinement, keyboard sliders and locale parity. Precision tests fail against the original PR implementation, which writes PCM16 regardless of the requested bits. The same implementation is used on macOS, Windows and Linux; real synthesis on macOS/Linux and engines absent from this Windows installation is not claimed here.
For another installed engine or reference, run the repository Python interpreter:
python scripts/compare_generation_quality.py \
--engines omnivoice omnivoice-subprocess voxcpm2 kittentts \
--output /path/to/new-comparison-directory \
--ref-audio /path/to/reference.wav --ref-text "Reference transcript"The script runs offline, uses installed models, routes WAVs through provenance marking when enabled and available, and writes
results.json, and refuses to overwrite an output directory. Omit the reference
arguments for TTS only; use --text, --steps and --seed to vary the experiment.
Keep private recordings and generated voice comparisons outside Git. The original
audio investigation remains available as historical
evidence; this work does not establish a cause for the historical reverb report.
Generated WAV reads use a shared TorchCodec fallback, including dub assembly, cached segments, watermark detection, longform and persona previews. Voice Clone blocks models that explicitly report no cloning support and explains how to select a capable model; plain text-to-speech remains available.