speech_to_text, diarize_speech, translate_text and summarize_text each repeat that warning; list_asr_models is the tool you can call freely to see what you would be paying for. The audio side works on local paths and writes its result back to disk rather than returning it inline, so a transcript is a file you then have to open — worth knowing before wiring this into anything that expects text in the response. translate_text and summarize_text take text directly, which makes them usable on their own and not only on what the transcriber just produced.
A Whissle client for spoken and written language: transcription, speaker diarization, translation and summarisation, plus a listing of the ASR models available to you.
- speech_to_text transcribes the file at audio_file_path using the ASR model named in model_name, and can return timestamps.
- boosted_lm_words and boosted_lm_score bias recognition toward terms you supply, so names and domain jargon survive into the transcript.
- diarize_speech produces the same transcript with speaker labels attached, bounded by max_speakers.
- Transcripts are written out as a text file to a directory you name, or to $HOME/Desktop when you name none.
- translate_text moves text between language codes given as source_language and target_language, such as en to es.
- summarize_text condenses the text in content using the model named in model_name, following any instruction you attach.
- list_asr_models reports the available ASR models and what each can do.
A Whissle auth token, supplied as WHISSLE_AUTH_TOKEN, together with UV_PYTHON set in the environment. Audio has to be a file the server can open at the path you pass as audio_file_path.