Optional splitter used only when the audio exceeds Groq's upload ceiling. See AudioSplitter for why this is injected rather than bundled.
ProtectedapiOnly sub-classes can access the API key
Read only getter for the underlying Groq client instance
Speaks params.text in params.voice. On success the audio is in SpeechResult.data, and
base 64 encoded in SpeechResult.content.
The driver's speech models.
The pronunciation dictionaries available to the driver; empty when it has none.
The names of the methods this driver implements.
The voices the driver can speak in.
Transcribes audio to text using one of Groq's Whisper models.
Pieces of split audio are transcribed sequentially, not in parallel: Groq rate limits by audio-seconds per minute, so firing an hour of audio at once buys nothing but 429s, and a partial failure mid-way would leave a transcript with an unmarked hole in it.
params.CancellationToken is passed to each piece's request, and no further piece is sent once
it aborts; the failure is then classified as a cancellation.
ProtectedTranscribeTranscribes audio that may exceed the provider's upload ceiling, splitting it first when it does. See TranscribeAudioWithSplitting, which this calls, for the behavior.
The full audio to transcribe
The provider's hard upload ceiling
Target piece size, below the ceiling to leave room for multipart framing
Splitter to use for oversized audio; may be null, in which case oversized audio fails
Provider name, used in the size-limit error messages
Transcribes a single piece already known to be within the ceiling, and is
handed cancellationToken to forward to its request
OptionalcancellationToken: AbortSignalCancels the transcription: no further piece is sent once it aborts
Groq implementation of BaseAudioGenerator, covering speech-to-text via Whisper.
Groq offers no text-to-speech surface, so
CreateSpeechand the voice/dictionary methods are not supported —GetSupportedMethodsreports exactly what works.Whisper on Groq does not do speaker diarization. There is no parameter for it and no way to emulate it from the response, so a caller that needs speaker labels needs a different provider (AssemblyAI, Deepgram) rather than a flag here.
Audio above Groq's 25MB upload ceiling requires an AudioSplitter, assigned to Splitter. Without one, oversized audio fails with a message naming the option rather than silently transcribing a truncated prefix.
Registered against BaseSpeechToText under the key model metadata names. It still extends, and is registered against,
BaseAudioGenerator, so callers that resolve it through the old base keep working.