Deepgram

Overview

Deepgram Voice Agent supports bring-your-own TTS through the agent.speak block of its Settings message. Its supported third-party types are open_ai, eleven_labs, cartesia, and aws_polly — there is no “custom Deepgram-dialect endpoint” option.

Speechify fits the open_ai slot. Point Deepgram at the tts-shims openai binary, and Deepgram sends an OpenAI-shaped POST /v1/audio/speech. The shim translates that request to a Speechify call and returns audio in the format Deepgram expects. Your Speechify key lives only in the shim’s environment.

Prerequisites

• Speechify API key
• Deepgram key with Voice Agent access
• Public HTTPS URL for the shim (Deepgram calls it from its cloud)
• Go toolchain locally (only to build the shim)

Build and run the shim

git clone https://github.com/Speechify-AI/tts-shims
cd tts-shims
make openai

Start it with your Speechify key:

export SPEECHIFY_API_KEY=sk_your_key_here
export SHIM_ADDR=:8771
./bin/openai

The openai binary mounts POST /v1/audio/speech (the OpenAI TTS endpoint shape). Confirm it’s up:

curl -s -o /dev/null -w "%{http_code}\n" http://localhost:8771/healthz

Verify without Deepgram

Send Deepgram’s exact request shape yourself:

curl -s -o shim-smoke.mp3 -w "%{http_code} %{size_download}\n" \
  -X POST "http://localhost:8771/v1/audio/speech" \
  -H "Authorization: Bearer placeholder-ignored-by-shim" \
  -H "Content-Type: application/json" \
  -d '{"model":"simba-3.2","input":"Deepgram voice agent, now speaking with Speechify.","voice":"geffen_32","response_format":"mp3"}'

A working shim returns:

200 <byte-count>

Inspect the output:

file shim-smoke.mp3
# MPEG ADTS, layer III, v2, 128 kbps, 24 kHz, Monaural

The Authorization: Bearer value here is a placeholder — the shim ignores it and uses SPEECHIFY_API_KEY from its own environment for the upstream call.

Configure Deepgram Voice Agent

In the Deepgram Voice Agent Settings message:

{
  "type": "Settings",
  "audio": {
    "output": { "encoding": "linear16", "sample_rate": 24000 }
  },
  "agent": {
    "speak": {
      "provider": {
        "type": "open_ai",
        "model": "simba-3.2",
        "voice": "geffen_32"
      },
      "endpoint": {
        "url": "https://your-shim-host.example/v1/audio/speech",
        "headers": {
          "authorization": "Bearer placeholder-ignored-by-shim"
        }
      }
    }
  }
}

The model and voice you set are forwarded verbatim to the shim, which passes them to Speechify. Set model: "simba-3.2" and pick any voice whose models list includes simba-3.2 (query GET /v1/voices to filter).

INVALID_SETTINGS: OpenAI speak provider type requires audio output format to be uncontainerized 24kHz linear16

Alias fallback (default voice)

If you leave the open_ai defaults (model: "tts-1", voice: "alloy"), the shim falls back to its own default model (SHIM_DEFAULT_MODEL) and the george voice. Set SHIM_DEFAULT_MODEL=simba-3.0 if you have not already - earlier builds defaulted to simba-english, which is retired from API version 2026-09-21 and returns 400 model_retired. Better still, set explicit model: "simba-3.2" and a Simba 3.2 voice in the Deepgram config for the current quality tier.

Production deployment

• Deploy the shim behind HTTPS on a service that holds environment variables.
• Set SPEECHIFY_API_KEY, SHIM_ADDR, and optionally SPEECHIFY_BASE_URL, SHIM_DEFAULT_MODEL, SHIM_REQUEST_TIMEOUT, SHIM_SHUTDOWN_TIMEOUT.
• Keep the placeholder Bearer token in the Deepgram config — it’s not a secret. The real key never leaves the shim host.
• Docker: docker build --build-arg PROVIDER=openai -t tts-shim-openai . (the repo’s Dockerfile takes the provider as a build arg; distroless runtime, static Go binary).

Resources