# /v1/audio/transcriptions + diarization/speaker endpoints
torchaudio
transformers>=4.35.0
faster-whisper>=0.10.0
openai-whisper>=20231117
vosk>=0.3.45
soundfile>=0.12.0
scipy>=1.11.0

# Speaker recognition — enrol/identify/verify voiceprints — is the `speaker`
# capability, which is served by THIS image. It shipped nothing for it: a
# speaker pod could boot and had no way to embed a voice. speechbrain imports
# fine on torch 2.11, so it needs no venv of its own.
#
# pyannote diarization is the part that genuinely cannot live here: it calls
# torchaudio.AudioMetaData, removed in torchaudio 2.11. That goes in the
# separate speaker image.
speechbrain>=1.0.0
