Wispr Building In-House ASR

Building its own speech layer turns Wispr from a thin wrapper around other companies APIs into a product that can keep more gross margin and get better with use. In practice, every dictated sentence otherwise carries a metered transcription cost to an outside model vendor. Owning ASR means Wispr can lower that variable cost, tune accuracy for messy real world dictation, and personalize to a user’s vocabulary over time.

• The economic logic is straightforward. Speech products process huge audio volumes, so even low per minute API fees compound quickly. Deepgram sells speech to text by usage, and OpenAI also prices transcription as a metered service. Replacing that with an internal model can materially lift contribution margin as usage scales.
  2 netlify 3 openai
• The product logic matters just as much as the cost logic. Wispr reports 10% word error rate versus 27% for Whisper and 47% for Apple dictation, which suggests its model is tuned for live dictation rather than generic transcription. Better accuracy means fewer manual fixes, which is what makes users trust voice input in email, docs, and messaging.
  1 sacra
• There is precedent for this move. Otter built a proprietary ASR engine to improve transcription quality and speaker separation, while Deepgram built its entire business on speech models sold by the minute through APIs. In speech software, the company that owns the recognition layer usually captures more of the value than the one simply reselling it.
  4 sacra 5 sacra

The next step is using proprietary ASR as the base for a broader voice stack. Once Wispr controls recognition quality and unit costs, it can push further into enterprise vocabulary, team dictionaries, and API products, which makes the software harder to replace and gives it more room to expand from a consumer dictation app into voice infrastructure.