The job it does
A voice agent has to know two things from the audio: what the caller said, and when they stopped. Deepgram answers both. My server opened one streaming connection to Nova-3 per call and forwarded the raw phone audio from Twilio. Deepgram sent back words as they formed, then a signal that the caller had finished a thought. That signal is what starts the model. Send it too early and the agent interrupts. Too late and the line feels dead.
I used the same setup in three places: the phone receptionist, a "talk to the assistant" button on its landing page, and voice demos I built for two sales pitches in June 2026.
Four settings to get right
The end-of-speech timer has a floor. utterance_end_ms would not go below 1000 when I built this. So the agent always waits at least a second of silence before it treats a sentence as done. I covered that second with a short filler phrase, cached as audio, that plays about half a second in.
Nova-3 does not take keywords. Older Deepgram models used that parameter to boost names and jargon. Nova-3 has no such parameter. Read the model's own parameter list before you copy a config forward.
Tell it the exact audio format. Phone audio from Twilio is 8 kHz mu-law. Browser audio is whatever the device's sound card runs at. Deepgram needs the encoding and sample rate stated for raw streams, and a wrong guess gives you silence or nonsense with no error.
Greeting echo. The agent's own greeting came back through the caller's line and got transcribed as if the caller had said it. A flag that ignores transcripts until the greeting finishes fixed it.
The Safari bug
The browser demo captured the microphone with MediaRecorder and streamed chunks to my server, which relayed them to Deepgram. In Chrome it worked. In Safari the assistant said hello and then never heard a word.
Chrome's MediaRecorder emits WebM, which decodes fine in pieces. Safari's emits MP4. An MP4 file keeps the index a decoder needs in a block that is written when recording stops. A live stream never stops, so the block never arrives, and Deepgram had nothing it could decode.
The fix, shipped June 4, 2026, was to stop using MediaRecorder. The page now reads raw samples through the Web Audio API, converts them to 16-bit PCM and sends those. Before the first audio frame it sends one message with the device's sample rate, and the server opens the Deepgram connection with linear16 at that rate. It works the same in every browser because no container format is involved.
Two Safari audio notes from the same week: close and recreate the AudioContext at the start of each session, because a reused one played silence with no error. And schedule the first chunk of speech about 120 milliseconds ahead, or the first word gets clipped.
One self-inflicted outage
On June 2, 2026 a rewrite of the server's environment file dropped the Deepgram and Cartesia keys. The phone line and the web demo both went mute. Nothing crashed, which is why it was easy to miss. A voice service should refuse to start when one of its vendor keys is missing. Mine did not, and it does not get a second chance to teach me that.
Prices on November 14, 2026
| Product | Pay as you go | Growth plan |
|---|---|---|
| Nova-3 streaming, one language | $0.0048 per minute | $0.0042 |
| Nova-3 streaming, multilingual | $0.0058 per minute | $0.0050 |
| Nova-3 pre-recorded, one language | $0.0043 per minute | $0.0036 |
| Flux streaming, English | $0.0065 per minute | $0.0057 |
| Voice Agent API, standard | $0.075 per minute | $0.068 |
| Voice Agent API, your own LLM and TTS | $0.050 per minute | $0.041 |
New accounts get $200 in credit that does not expire. Pay-as-you-go allows 150 streaming connections at once, which is 150 simultaneous calls.
When I priced my receptionist in spring 2026 I budgeted $0.0077 a minute for Nova-3. Today's rate is lower. At $0.0048, a thousand minutes of calls costs $4.80 to transcribe.