DP ref encoder uses compressed latent channels from config; README v2 onnx path; pytest smoke tests

Hebrew dates: spell month with ordinal names (לראשון … לשנים עשר)

Hebrew: format expanded dates as 'D ל-M Y'

Expand DD/MM/YYYY-style dates before generic number expansion

Hebrew: expand 'ל' before Latin tokens (e.g. CPU-GPU) to 'אל'

Lower BLUE_SYNTH_MAX_CHUNK_LEN to 200

Add phonemes_for_display and strip language tags for user-facing text

Strip Hebrew gershayim/quote marks in abbreviations before synthesis

Hebrew+Latin split phonemization; email/%/ratio/number prep; per-chunk phonemize; mixed pace blend

Tuning: CFG 4.0; tighten steps (5-16) and speed (0.8-1.2) slider ranges

Strip Hebrew nikud; normalize 'any more'; default speed 0.95, CFG 3.5

Download default female/male v2 voice JSONs from hub; restore style_ttl std filter; remove Refresh voices UI

Voices: require style_ttl+style_dp; add Refresh voices button; select new clone in dropdown

Gradio: drop CFG/pace sliders (use cfg 3, pace blend 0); fix bad indents in infer loop and helpers

Clone export: reference audio cleanup and left-aligned ref mask; PT weights bundle marker and force download

Strip XML-like tags with attributes; unify tag stripping for phonemize and char encode

Strip lang_list helper markup; restrict inline language tags in regex

download_models: ONNX bundle stamp to bust Hub file cache on rebuild

Restore vocoder edge trim after each chunk

Pace blend for duration model; long-text Gradio slider; strip inline lang tags in phonemize

Update README.md

Update README.md

Update README.md

Restore vocoder edge trim before waveform return

Post-process TTS: gentle RMS boost; drop vocoder edge trim

Centralize max synth chunk length in BLUE_SYNTH_MAX_CHUNK_LEN

Align CFG slider range with model; clean torchaudio imports in export

Resample reference WAVs with torchaudio for voice export

Improve voice clone UX with cache and Gradio progress

Simplify Gradio voice cloning and auto-run synthesis

Use v2 model repositories for Space runtime

Simplify uploaded voice style export

Force uploaded voice extraction to use PT checkpoints

Return uploaded WAV voice styles directly from exporter

Use root tts config for voice export normalizers

Make uploaded reference audio an explicit voice source

Align voice cloning path with reference exporter

Prefer pt_models (.pt) checkpoints matching vendored architectures

Infer DPNetwork vocab_size from checkpoint embedding

Extract style_dp from reference WAV via DPNetwork.ref_encoder

Clamp DP duration to text-length-based cap (fix runaway cloned voice duration)

Load vocab.json from repo root (fix noise output)

Auto-detect vocoder input name (z_pred vs latent)

Clamp predicted latent length below vector_estimator's ~1000-frame cap

Hard-split oversize chunks; lower max_len below vector_estimator's 1000-token cap

Voice clone: legacy state-dict remap + better error surfacing

Install bluecodec package so LatentEncoder is importable

Inline voice cloning: reference WAV directly on Synthesize tab (no JSON)

Add torch/torchaudio/librosa/safetensors for voice cloning

Skip DPNetwork in export; fall back to default-voice style_dp at runtime