Version 1.3.0 - Auto-Split Audio, Dataset Management & Engine Controls

[FranckyB]

Version 1.3.0 - Auto-Split Audio, Dataset Management & Engine Controls

Auto-Split Audio

• Automatic Audio Splitting - Split long audio files into clean sentence-level clips using Qwen3 or Whisper's timestamp extraction.
• One-Click Dataset Creation - Split audio and auto-save segments with transcripts directly into dataset folders
• Trim and discard Silent areas - Uses the timestamp data to find and remove non verbal moments.

Unified ASR Engine

• Unified ASR Dropdown - Single dropdown for all transcription engines (Qwen3 ASR, VibeVoice ASR, Whisper) replacing the old radio + size selector
• ASR Engine Toggles - Enable or disable individual ASR engines in Settings, just like TTS engines
• Dynamic Defaults - ASR dropdown automatically picks the best available engine based on what's installed and enabled
• Added Whisper Large - With the addition of automatic disply of available ASR engine, adding more choices doesn't bloat the ui.

Engine Availability Checker

• Startup Engine Detection - App now auto-checks which TTS and ASR engines are installed at launch
• Auto-Disable Missing Engines - Engines that aren't installed are automatically hidden from dropdowns
• Clean Console Output - Clear status report showing which engines are available, skipped, or missing

Dataset Management

• Create Dataset Folders - Create new dataset folders directly from the Prep Audio UI
• Manage Existing Datasets - Delete dataset folders with confirmation modal
• Drag & Drop Audio - Import audio files by dragging them into the editor

Quality of Life

• Renamed "Qwen CustomVoice" to "Qwen Speakers" - Clearer label in Conversation and Voice Presets tabs

• Overwrite Protection - Inline confirmation bar when saving a file that already exists

• Friendly Port Error - Clean message when Voice Clone Studio is already running instead of a traceback

• Whisper Now Optional - Moved from auto-install to optional in setup wizard, same as Qwen3-ASR.

• Suppressed Noisy Warnings - Silenced verbose k2 and flash-attn warnings during engine checks

• Qwen3 ASR Integration - Added Qwen3-ASR as a new transcription engine in Prep Audio, supporting 52 languages and dialects

• Model Size Selector - Choose between Small (0.6B, fast) and Large (1.7B, best accuracy) Qwen3 ASR models

• Language Selection - Qwen3 ASR supports language hints for improved accuracy, shared with Whisper's language dropdown

• ICL (In-Context Learning) for Trained Models - Enhanced Voice Presets with optional ICL mode that provides real-time prosody and style cues on top of trained voice identity

• Dataset-Based ICL Samples - Select reference audio from your training datasets for ICL, with audio preview and automatic transcript loading

• Speaker Encoder Transplant - Automatic fix for trained model checkpoints missing speaker encoder weights, loading them from the matching base model at runtime

• Setup Script Integration - Qwen3 ASR offered as optional install in setup-windows.bat, setup-linux.sh, and Dockerfile

• Suppressed Gradio HTTP Logs - Silenced noisy httpx/httpcore info-level logs from Gradio 6

---

This discussion was created from the release Automatic Dataset creation with addition of Auto Clip splitting.