fix: restaurar frontmatter YAML do HF Space no README

docs: README completo com API, arquitetura e integração AVA

perf: sentence-splitting para síntese por sentença dentro de uma GPU call

perf: pré-processamento de texto, cache de voz e max_new_tokens dinâmico

feat: adicionar transcrição tts-hub-pt e definir como voz ativa

add voice: tts-hub-pt sample

set active voice: ava-thalita

add transcript: ava-thalita

add voice: ava-thalita

cleanup: remove debug prints, fix ui_generate error handling

fix: collect audio from code='final' (non-streaming mode returns full audio there, not in 'segment' results)

fix: torchaudio.load stub must handle raw bytes (wrap in BytesIO), add load diagnostics

debug: log all inference result codes to diagnose empty audio

debug: return RuntimeError traceback as bytes to surface actual GPU error

refactor: lazy-load models on GPU inside @spaces.GPU(duration=180) on first synthesis call

fix: @spaces.GPU(duration=300), force device=cpu to match LLM thread (avoid mismatch)

fix: add torchaudio.functional.melscale_fbanks to stub (uses librosa.filters.mel)

fix: remove torchaudio from requirements (CUDA ABI conflict), stub transforms.MelScale via librosa

fix: ModuleSpec is in importlib.machinery not importlib.util

fix: set __spec__ on torchaudio stub so importlib.find_spec doesn't raise ValueError

fix: stub torchaudio entirely when CUDA ABI mismatch on ZeroGPU (use soundfile+librosa)

fix: torchaudio.load stub must pass BytesIO directly to soundfile, not str()

fix: replace torchaudio.load with soundfile (torchcodec not available on HF)

feat: upload ava-ejae reference voice (via HF API, Xet storage)

feat: add ava-ejae voice config and transcript (wav via HF API)

fix: use cuda only if available, fall back to cpu for synthesis

fix: monkeypatch torchaudio.list_audio_backends (missing in old torchaudio)

fix: RankedLogger stub must delegate to loguru (needs .info/.debug/etc)

fix: add all missing fish-speech inference deps at once (librosa, resampy, einx, rich, etc)

fix: add loralib dep (fish_speech.models.text2semantic.lora)

fix: add transformers and accelerate (needed by text2semantic inference)

fix: stub lightning_utilities and fish_speech.utils.logger

fix: don't stub fish_speech.utils package, only stub lightning submodules

fix: set __file__=None on stubs, raise AttributeError for dunders

fix: __getattr__ in module stubs takes (k) not (self, k)

fix: stub lightning/pytorch-lightning training imports in fish_speech.utils

fix: add pytorch-lightning dep for fish_speech.utils import

fix: use fish-speech@v1.5.1 with correct API and all required deps

fix: install fish-speech --no-deps to avoid gradio version conflict

feat: initial Fish Speech 1.5 TTS Space

initial commit