Deepgram on the AI Gateway: STT, TTS, Cost and Governance

What Deepgram is

Deepgram is a speech-to-text and text-to-speech provider built for real-time, production voice. Its current flagship transcription model, Nova-3, handles both pre-recorded and streaming audio, with smart formatting, diarization, and multilingual support. On the output side, Aura turns text into natural speech for voice agents and IVR.

Within any transaction of the voice app, they occupy opposite sides; the caller says something, then Nova-3 converts the audio signal into text, an LLM infers from that information, and then Aura (or other TTS engine) says the result aloud. With Deepgram being quick and accurate, it is becoming the default tool for both transcription hops and increasingly speech hops.

The problem lies in the practice. Deepgram is yet another vendor, with yet another API key, yet another SDK, yet another expense — and since it is billed based on the time and number of characters transcribed and not by tokens, it is not part of your spending tracking of LLM. But for one design it will be okay. In an enterprise with a team of voice bots, it’s a case of governance failure.

Deepgram pricing

Here is the pay-as-you-go rate card for the models you are most likely to use.

Two things to note. First, the unit is not the token. STT is billed per minute of audio, TTS per 1,000 characters of input text. That mismatch is exactly why voice cost tends to escape token-based dashboards. Second, streaming costs more than batch — real-time transcription runs higher than pre-recorded, so a live voice agent is priced differently from an offline transcription pipeline. New accounts get $200 in free credit to start [VERIFY current rates and credit at deepgram.com/pricing; promotional streaming prices have fluctuated].

Why run Deepgram through a gateway

A voice agent is not one model call — it is a chain, and every hop is a provider you pay, monitor, and can have fail on you. Putting the whole chain behind one gateway turns that sprawl into a single governed path.

A voice agent through the TrueFoundry AI Gateway

Routing Deepgram through the TrueFoundry AI Gateway gives you four things you do not get calling it directly:

• One API key and endpoint for STT, TTS, and the LLM in between. Your app holds a TrueFoundry key; the gateway holds the Deepgram key. Rotating or revoking Deepgram access is a gateway change, not a redeploy.
• Unified cost tracking that includes audio. The gateway meters Deepgram usage per minute and per character and shows it alongside your LLM spend, attributed by team, user, and request.
• Fallback and load-balancing across STT or TTS providers. If Deepgram is rate-limited or degraded, a virtual model can fail over to another transcription provider without a code change.
• Access control and guardrails applied uniformly — the same collaborator model, the same policies, across every provider.

STT and TTS, one endpoint

Deepgram is not the only speech provider the gateway supports, which is what makes fallback and comparison possible.

Speech-to-text support across providers in the gateway

Deepgram, Cartesia, ElevenLabs, Google Vertex, and Smallest AI are called through the gateway’s provider proxy using each one’s native SDK, while OpenAI, Azure, and Groq use the OpenAI-compatible API. That means you keep Deepgram’s exact feature set — you are not flattening it to a lowest-common-denominator API — while still getting one key, one bill, and one failover layer.

Where teams get this wrong

They track LLM cost and forget audio. Token dashboards miss per-minute and per-character spend entirely. On a busy voice product, transcription and speech can rival the LLM bill — and nobody sees it until finance asks.

They hardcode the Deepgram key in each service. Every voice service holding its own Deepgram credential means rotation is a redeploy and there is no central place to cap or attribute usage.

They have no failover for the voice hops. Teams build LLM fallback and leave STT/TTS as single points of failure. If Deepgram has a bad minute, the whole agent goes mute.

They confuse streaming and batch pricing. Running a real-time agent on the assumption of batch rates understates cost. Streaming is priced higher; model it that way.

‍