Audio & Transcription

Kairos transcribes whatever audio source it's listening to. The quality of your detections depends on the quality of that transcript, which depends on two things: the microphone you point at, and the speech provider you choose to transcribe it with.

Picking an input device

Open Settings → Capture → Audio to choose your microphone. Every audio input the operating system sees shows up in the list, the built-in mic, any USB audio interface plugged in, a USB lavalier, or a virtual device routed in from your soundboard.

The live input meter next to the picker has a Test button that opens the mic and shows real-time RMS and peak levels (with a brief peak-hold so transients register). A healthy level peaks in the upper third of the meter when the pastor is speaking, and sits near the bottom in silence.

Choosing a speech provider

Kairos supports four transcription providers. Switch between them in Settings → Capture → Speech.

Local, Whisper

Runs Distil-Whisper large-v3 locally on your machine. Audio never leaves your network.

• Pros: fully offline, no API key, no per-minute cost, audio stays on your device.
• Cons: uses real CPU/GPU resources. On lower-spec machines it can fall behind a fast preacher.
• When to pick it: your venue has unreliable internet, your privacy policy requires audio to stay local, or your hardware can handle it.

Cloud, Deepgram

Streams audio to Deepgram Nova-3 over a WebSocket. Best accuracy of the cloud options, with Bible keyword boosting baked in.

• Pros: fast, accurate, very low latency. Bible-aware.
• Cons: needs internet, sends audio to Deepgram, costs per minute of audio.
• When to pick it: your venue has solid internet, you want the highest accuracy, and you're comfortable with the cloud cost.

Paste your API key into the Deepgram API Key field after selecting this provider. Get one from deepgram.com.

Cloud, OpenAI

Posts each utterance to OpenAI's transcription API in batches. Three model choices, with a noticeable price/quality tradeoff:

• Pros: cheapest cloud option (with gpt-4o-mini-transcribe).
• Cons: batch mode adds ~1–3s latency per utterance.
• When to pick it: you want cloud accuracy but Deepgram and Gladia are out of budget, and the extra latency is acceptable for your service style.

Paste your API key after selecting this provider. Get one at platform.openai.com.

Cloud, Gladia

Real-time streaming via Gladia's v2 live API.

• Pros: low-latency partials, code-switching available on higher tiers (useful for bilingual services).
• Cons: needs internet, costs per minute.
• When to pick it: your service is bilingual or you want a Deepgram alternative with similar latency characteristics.

Paste your API key after selecting this provider. Get one at app.gladia.io.

Tuning the gain

If the live input meter peaks too quietly (you only see a sliver of movement) or too aggressively (it pegs at the top constantly), adjust the gain slider in Settings → Capture → Audio. Aim for peaks in the upper third when the pastor is at normal volume.

A clean signal at the right level beats any amount of detection tuning. Get this part right first.

Troubleshooting

• Transcript is empty: wrong input device selected, or the OS hasn't granted Kairos microphone permission. See Troubleshooting.
• Lots of garbled words: likely a noisy room mic. Try switching to a soundboard feed or moving to a closer mic.
• Transcript is delayed by several seconds: if you're on local Whisper, your hardware may be falling behind. Switch to a cloud provider (Deepgram or Gladia for streaming, OpenAI for batch). If you're on OpenAI, that ~1–3s delay is expected, switch to Deepgram or Gladia for streaming.