Clarify emotion field scope across dataset workflows Dataset Builder's "Emotion" column is a VoiceDesign generation prompt, not training metadata. Users preparing real audio ZIPs were confused about whether they need an emotion field in metadata.jsonl (they don't). Added cross-references between the two pages to prevent confusion. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update docs: EPUB import, training language, configurable pauses, remove sub_batch_max_chars
Updated Training Guide (markdown)
Fix metadata.jsonl docs: remove instruct field (matches Qwen3-TTS dataset spec) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add ref audio instruct tip for LoRA training pacing
Document LoRA rank selection: r=32 vs r=64 tradeoffs
Add Dataset Builder page, update wiki for new features - New Dataset Builder wiki page with full workflow docs - Add Dataset Builder API endpoints to API Reference - Remove stale generation config field from API Reference - Add sub_batch_max_chars to Batch Generation settings table - Add Dataset Builder as Method 3 in Training Guide - Mention built-in LoRA presets in Training Guide and Voice Types - Fix saved scripts directory name in Script Generation - Add Dataset Builder to Home navigation Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update training docs: 4.1 loss floor, Watson series data, run-to-run variance
Add sample duration mix guidelines to training guide
Fix ref sample docs: configurable dropdown, not first sample
Update Training Guide with empirical data and add Voice Reference page
Add wiki documentation: Voice Types, Training Guide, Batch Generation, Script Generation, Editor & Export, API Reference, Troubleshooting