Vox Builder

s3g Vox Builder is a macOS companion app for creating voicebanks used by Ambi Encoder Vox. It can synthesize a reproducible voice source or turn one continuous vocal recording, several WAV files, or a folder of WAVs into an editable set of aliases without embedding WORLD data in the exported bank.

Workflow

1. Open Vox Builder, then generate a voice source, record or load a continuous WAV, or load or drop WAV files or a folder.
2. Select a segment and click AUDITION to establish source monitoring before editing.
3. Set the alias inventory and run ANALYZE to segment the source and estimate its voice metadata.
4. Audition each segment, correct its alias and timing, and adjust boundaries or analysis settings as needed.
5. Click EXPORT VOICEBANK, keep the complete exported folder together, and load that folder in Ambi Encoder Vox.

Voice Generation

GENERATE opens the Voice Source Generator. It creates a dry mono source at 48 kHz using a procedural source-filter voice model with voiced excitation, breath and turbulence, formant motion, consonant bursts, and alias-specific articulation. CORE 35 generates the compact five-vowel by seven-onset recording matrix; FULL 92 generates every CV alias used by the built-in English phrase mapper.

PROCEDURAL starts from the generator controls alone. SEEDED first measures the currently selected segment's pitch, broad formant balance, breath, roughness, brightness, and periodicity, then uses those measurements to condition a newly synthesized bank. It does not copy or montage source samples. Seeded mode therefore preserves broad vocal character while still allowing substantial variation and a fully generated result.

PITCH and TRACT establish the voice range and resonant scale. BREATH, ROUGH, and VARIATION control excitation character and alias-to-alias difference. CLARITY, CONSONANT, and DURATION set articulation strength and timing. The integer SEED makes the result repeatable: identical parameters and seed produce identical samples.

The generated aliases enter the same waveform, audition, edit, and export workflow as recorded material. voicebank.json records the generation mode, vocabulary, controls, random seed, and any measured seed profile so a bank's origin can be inspected or recreated.

Source Recording

For a continuous source, record the aliases in a known order and leave a short quiet gap between syllables. Load or drop the WAV, then edit the alias list so it matches the recording. The included example uses examples/voicebank-builder/phonemes.txt and s3g_demo_voice_source.wav.

LOAD WAV / FOLDER also accepts multiple selected WAVs or a folder. Folders are scanned recursively, files are naturally sorted by path, and each WAV becomes one exact segment with its filename as the initial alias. All files in one import must use the same sample rate. Multiple files and folders can also be dropped onto the waveform area.

Analysis

Before segmentation or WORLD analysis, Vox Builder removes DC offset and normalizes active RMS toward -18 dBFS while limiting peaks to -3 dBFS. Gain boost is limited to 24 dB so a nearly silent or noise-only recording is not made deceptively loud. Multi-file imports are conditioned per file for more consistent alias levels. The Source panel reports the applied gain range and flags source clipping.

Empty audio, recordings shorter than 50 ms, and mixed sample rates in one file set are rejected before the current bank is replaced. A clipped recording can still be inspected and exported, but remains marked for review because reducing its level cannot reconstruct clipped peaks.

For a continuous source, THRESH sets the active-audio threshold, MIN GAP sets the quiet duration required between regions, and PAD retains context around each detected region. ANALYZE rebuilds the segments after any of these controls or the alias order changes.

For a multi-file source, filenames are first matched against the editable alias inventory. Unmatched files receive the next unused alias in natural file order, then fall back to their filename. The Segment readout shows M as match confidence; editing an alias marks that assignment as confirmed.

ANALYZE applies the edited alias list before repeating WORLD analysis. The alias count must match the current segment count.

When silence-separated regions do not match the alias count, the builder searches the amplitude envelope for low-energy boundaries while preserving one ordered segment per alias. WORLD Harvest estimates each segment's root MIDI note and voiced ratio, with DIO and StoneMask as fallback.

Alias Vocabulary and Guessing

Ambi Vox accepts any clean exact alias. A custom alias can be typed directly in a phrase or reached through s3g-pronunciations.txt. The finite automatic vocabulary is the set of CV tokens emitted by the built-in English phrase mapper. Open ALIAS GUIDE in Vox Builder to see that complete vocabulary, its fallback pairs, and the core recording matrix. The Segment ALIAS field is an editable dropdown containing every automatic phrase token while still allowing a custom value.

AUTO GUESS assigns uncertain segments from acoustic analysis, using LPC formant estimates for the vowel and onset energy, periodicity, high-frequency content, and transient shape for the consonant class. The acoustic classifier intentionally targets the core 35-alias matrix: a/e/i/o/u and the k/s/t/n/m/r CV rows. Each core alias is assigned at most once per pass: when the best match is already used, the segment receives its next available candidate even when that match has low confidence. Exact filename matches, manually edited aliases, and a high-confidence recording order are retained and reserve their aliases. After assignment the button becomes UNDO GUESS, and low M values should be auditioned and corrected.

This is a fast labeling estimate, not general phoneme recognition. Exact distinctions among a larger consonant vocabulary require a trained speaker-specific model, so Vox Builder does not report unsupported certainty. Filename labels, a known recording order, and manual confirmation remain the reliable routes for the complete automatic phrase vocabulary.

Editing

Click a numbered region to select it. Drag an interior vertical boundary to adjust two neighboring regions together. Detected, split, and dragged interior boundaries snap to the nearest zero crossing within 10 ms while preserving a minimum segment length. The Segment header + splits the selected region near its midpoint and - removes it by merging it with a neighbor. The new boundary remains draggable, and the alias list stays synchronized.

AUDITION plays the selected slice. Alias, root MIDI, fixed/consonant duration, preutterance, and overlap can be entered directly in the Segment panel. The readout also shows voiced percentage, alias-match confidence, duration, and applied normalization gain.

The root and voicing readout is diagnostic metadata. Ambi Vox performs its complete CheapTrick, D4C, pitch-anchor, and cache analysis when the exported bank is loaded.

Export

EXPORT VOICEBANK creates a folder containing one WAV per alias plus oto.ini, character.txt, phonemes.txt, markers.csv, and voicebank.json. Load that folder from Ambi Encoder Vox.