Native STT vs. Deepgram: The Cost-Quality Trade-off

TAMSIV's voice pipeline relies on Deepgram. Excellent quality, impeccable French support. The problem? Every second of audio has a price.

Native STT as an alternative

Every smartphone has a built-in speech recognition engine. Free, local, fast. Less accurate than Deepgram for fast French, no reliable punctuation. But for dictating a short task, it's sufficient.

The dual architecture

Two interchangeable modes, configurable from the admin:

• Cloud (Deepgram): Audio via WebSocket, integrated VAD, high quality. Pro/Team plans.
• Native: On-device recognition, no data leaves the phone. Free plan.

The frontend exposes a unified interface. Components don't know which engine is running.

Real-world comparison

• Quiet environment: native 92%, Deepgram 98%
• With background noise: native 75%, Deepgram 94%
• French with accents: Deepgram significantly better

The verdict: Deepgram is superior. But for a free plan, native is acceptable. Users who want the best quality have another reason to go Pro. Win-win.