executor_latency_seconds: more granular buckets up to 1.5 hour of latency

ObservedExecutor to monitor too long workloads

hf_hub_download_hang_me_weirdly case for repro incidents

README.md: TTS & STT

tts: streaming with explicit streaming: true

encoding with soundfile simplified + opus/ogg support

final tunes before production

fix

kkr-requirements-cuda.txt

load_uz_dataset.py

openai like http exceptions formatting

load_uz_dataset.py

respond with 400 in case of bad lang,voice; if_sara for eo

lang ident not blocking/async way: in work thread

final lang and voice exposed via headers

auto for voice, lang

detect_to_espeak()

language identification via fasttext

cleaning GPU resources on pipeline unloading, because it also pushes tensors to GPU: voices

tts: unloading on idle

compat mode: language from voice as first letter

response_format other than wav (like mp3) streamable and not

fastapi.UploadFile: sync vs async

/v1/audio/speech: raw implementation

kkr/app service

fixed mypy issues