Add audio transcription support (Whisper) to OpenSecretClient

[elimisteve]

Summary

OpenSecretClient currently supports chat completions, embeddings, and model listing, but not /v1/audio/transcriptions. The Maple backend exposes whisper-large-v3 as an available model, but there's no way to reach it through the SDK.

Problem

encrypted_openai_call is private and only supports JSON-serialized request bodies. Audio transcription requires multipart/form-data with a file upload. Because session_manager and encrypted_openai_call are private, downstream consumers (like maple-proxy) cannot implement transcription support themselves.

Requested API

pub async fn create_transcription(
    &self,
    file: Vec<u8>,
    filename: String,
    model: String,
    language: Option<String>,
    response_format: Option<String>,
) -> Result<TranscriptionResponse>

This would need to encrypt the audio file using the session key, send it to /v1/audio/transcriptions, and decrypt the response.

Context

Building a local proxy (maple-proxy) that exposes an OpenAI-compatible API on localhost. Chat completions and embeddings work, but audio transcription is blocked by this missing SDK method.

[AnthonyRonning]

Yeah, been needing to expose this via API. It's mostly just that I'm not sure what/how to charge for it. I don't really want to do too much parsing of the file upload to get stats like length of recording. Might just make it based on how big the file is.

[elimisteve]

Yeah file size or duration (if cheap to parse; surely quite cheap compared to processing it!) or how long it takes to process/transcribe all make sense to me; any would do, but pegging it to your actual costs (GPU-seconds?) seems most sustainable on your end.