Expand description
Whisper speech-to-text on whisper.cpp.
Decoding follows what open-speech asks of faster-whisper: beam search of
width 5, temperature 0, an optional initial prompt, and a language that
is detected unless given. Input is 16 kHz mono f32
(loqui_audio::decode_mono_16k produces it from any upload).
Weights are GGML files such as ggml-large-v3-turbo.bin from
ggerganov/whisper.cpp; they are not bundled.