Use the same manual_seed for every chunk and normalize per-segment peaks before concat so long voiceovers no longer sound like different speakers between segments.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep synthesized wav files browsable with playback and download, default to preset steady male voice, show one-click pipeline as the first tab, and reduce post-synthesis UI flicker.
Co-authored-by: Cursor <cursoragent@cursor.com>
Enable Gradio queue, immediate pending feedback, segment progress, and gr.update for Audio so long syntheses show logs and playback correctly.
Co-authored-by: Cursor <cursoragent@cursor.com>
Re-enable KV cache by default, normalize digits and unsafe chars, disable per-chunk split_text, and reload ChatTTS after CUDA errors.
Co-authored-by: Cursor <cursoragent@cursor.com>
Disable ensure_non_empty retries, set min_new_token, always refine text, and use per-chunk manual_seed.
Co-authored-by: Cursor <cursoragent@cursor.com>
Release GPU memory before TTS/ASR switches, lower TTS token limits, and set PYTORCH_CUDA_ALLOC_CONF in PM2.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use spk_smp plus txt_smp for voice clone instead of mis-encoding into spk_emb; migrate legacy speaker_emb.pt and improve error hints.
Co-authored-by: Cursor <cursoragent@cursor.com>
Strip Markdown and stage directions before ChatTTS synthesis with chunked long scripts; document model pre-download, server-update, and microphone HTTPS notes.
Co-authored-by: Cursor <cursoragent@cursor.com>
dekun