Skip to main content

Larger custom dictionary, voice fine-tuning, and free system instruction for the custom API

Feature Request

The dictionary limit is the main thing holding the free tier back

I'm a Vowen user, and the 50-word dictionary cap is honestly the only free-tier restriction that gets in my way. I'd happily see a much larger dictionary as a paid feature - it would be genuinely valuable for me and several people around me. For reference, wisprflow.ai lets you add thousands of words and tunes the system around them (optimal LLM prompts, etc.).

I've set up a small model that runs locally on my laptop and proofreads my transcribed text. It comfortably handles far more than 50 words using the language model alone - I'd estimate 500 - 1000 words without trouble on my hardware, and I'd expect the same via API requests. So the technical ceiling is much higher than the current limit suggests.

The feature I'd really love: personal fine-tuning

What would be amazing is a way to fine-tune the model specifically for me - for my voice and my vocabulary. If there's any way I can help test or contribute to this, I'm in.

Feedback: please don't paywall the system instruction on the custom API

Locking the LLM system instruction behind a paywall doesn't make much sense. It's trivial to work around - with LM Studio, Llama, or any other provider, I can set a system instruction at the provider's own level when connecting the model. So a paid system instruction just pushes me to configure it elsewhere. Given that it used to be free, the better move is the opposite: keep the system instruction as a free feature.

Posted by Nikolai Vysotskyi ·
1 comment

Log in to comment and vote

Comments1

  • Vowen

    Team•

    May 25

    Hi @Nikolai Vysotskyi Thank you for the valuable feedback. Here are some clarifications regarding your questions:

    1. On the 50-word dictionary cap — it's not a Pro paywall, Pro hits the same limit. If I remember right, we landed at 50 because we got issues raised by users in the past that there was a clear degradation in transcription accuracy (the model starts hallucinating biased terms into the output). Different models enforce different hint limits, so 50 was decided for batch models and 100 for streaming models. That said, the limit shouldn't be uniform across models — local setups likely could have more headroom. We'll look into this issue.

    2. Personal fine-tuning — We don't retain voice data, so there's no per-user dataset to train against. An opt-in program where users explicitly volunteer recordings is probably the only path, but nothing's on the roadmap yet. Our original plan to build this was this exact reason, we wanted to build a product which performs just as well when given the right instructions. If we open a beta for training, we’ll surely get back on this.

    1. Custom instructions for AI text enhancement is a free feature. It was unintentionally locked when rolling out the paid plan and we’ve fixed it in our latest release. If you update to the latest version, the system instruction field should work on the custom API path (and everywhere else) regardless of plan. Please check and let us know if you’re seeing any issues.