Support OpenAI-Compatible Third-Party STT Providers

[akshara-tg]

Are there any plans to support other third-party speech-to-text service providers that are compatible with OpenAI?

For example, I have been using APIs from AssemblyAI to generate transcriptions for a significant portion of my audio data.

Do you have any intentions of supporting these third-party service providers that are compatible with the OpenAI speech-to-text endpoint?

[murtaza-nasir]

Does this not work with whatever service you're trying to use? It works with openAI, and it is supposed to work with any openAI compatible endpoint. You provide the endpoint URL in your configuration.

[razbradley1]

First, thank you for all the work on Speakr. It’s a really impressive project, and I’m excited about the possibility of using it for local/self-hosted meeting transcription.

I wanted to revisit this request and ask if you have considered adding AssemblyAI as a transcription option.

The main reason is that AssemblyAI seems to handle longer meetings better, especially meetings in the 1–3 hour range with 5–8 speakers. In my testing, it appears to do a better job with speaker diarization and keeping things usable across longer, multi-speaker conversations than the current OpenAI transcription options.

I know AssemblyAI is not a drop-in OpenAI-compatible endpoint, so this would probably require native support or some kind of provider abstraction. But even a basic AssemblyAI integration would be really helpful for people using Speakr for real meetings with multiple speakers.

Thanks again for building this!

[murtaza-nasir]

@razbradley1 AssemblyAI is now a built-in connector (on master, in the next release). Set TRANSCRIPTION_CONNECTOR=assemblyai and TRANSCRIPTION_API_KEY=<your key>. It runs AssemblyAI's own upload/poll flow, diarizes via speaker_labels, and handles long files in a single job (up to 10 hours, no chunking), which is the multi-hour multi-speaker case you described. Speakr's hotwords map to word_boost and min/max speakers to speaker_options. I have tried it with some basic testing. It will become available in the next release, or you can pull and build it yourself in the meantime.

[razbradley1]

@murtaza-nasir Incredible. Thank you so much!

[razbradley1]

@murtaza-nasir Can confirm it is working well. And I re-ran a previous longform transcription that took about 18 mins on OpenAI, and it returned from Assembly in good condition in about 45 seconds.

Very appreciative of this connector.