Harden production inference normalization and boundaries

Harden fail-closed inference selection

Recognize simplified frequency units

Reject ambiguous frequency collisions

Fail closed on ambiguous frequency forms

Verify normalized Chinese frequency targets

Normalize frequency proof sentence context

Canonicalize target-proven frequency scales

Pin default built-in speaker contract

Add coverage-adaptive chunk refill

Bound active pace correction to gate margin

Remove artificial onset chunk splits

Add active-voiced pace correction ablation

Harden target-proven numeric normalization

Add dual-ASR whole-output verification

Retain content-free cascade diagnostics

Stabilize verified sequence fallback

Make quality ASR decoding deterministic

Harden hosted inference verification

Add fail-closed stable speaker inference

Fix Space dependency resolution

Pin verified model and speaker encoder revisions

Catch weak stop signals near sentence end

Stabilize very short utterances

Prevent speech after the requested text

Stop cleanly at natural text boundaries

Stop earlier without clipping final speech

Stop forcing generated tail speech

Finish forced endpoints cleanly

Tighten onset pacing and endpoint cap

Tighten endpoint stop window

Stabilize onset pace and endpoint stopping

Fix SpeechBrain on ZeroGPU torchaudio

Use stable production inference profile

Replace examples with Taiwan-specific-reading test sentences (期/績/括/究/法/企/髮/說服/骰子)

Add more Taiwan-flavored example texts (小吃、機車/捷運/悠遊卡、語尾詞) + night-market stream example

Add example text: 我要吃蚵仔煎然後去丟垃圾。

Add speech-rate slider (pitch-preserving time-stretch post-processing, shared across tabs)

Add adjustable cfg_value / inference_timesteps sliders (shared across tabs, defaults = model-card recommendation)

Voice clone: single reference-clip mode only (drop centroid/ECAPA stable mode + speechbrain)

Wording: neutral phrasing for reference-clip support boundary

Update for checkpoint step_0006000

Update for checkpoint step_0006000

step_0006000: fast transcript-free clone (reference_wav_path), new default params (cfg 2.0/steps 10)

Add speaker selection (李宏毅 + 女聲) via multi-speaker centroids

Voice clone: read 10 sentences -> average ECAPA over all clips for a stabler centroid

Voice cloning: extract ECAPA speaker centroid from upload (no transcript); add speechbrain

Fix voice cloning: use prompt_wav+prompt_text (reference_wav_path produced gibberish); remove debug

TEMP: add /tts_debug endpoint to diagnose voice-clone content failure

Long-text streaming: stream until GPU time budget (~100s) instead of fixed 12-sentence cap