# Deepgram > Use Deepgram for streaming speech-to-text in a cascaded voice agent, with low-latency interim transcripts and barge-in. Deepgram is the default speech-to-text provider for a cascaded FirstByte Voice agent. When you run an ASR -> LLM -> TTS pipeline, the caller's audio streams into Deepgram over a persistent socket. Deepgram returns interim and final transcripts. The runtime sends the transcripts to the language model as soon as the caller stops talking. This integration is **shipped**. A bare cascade with no `asr_provider` set already uses Deepgram. You set one config field. You add a key. You place a call. You do not write code and you do not rebuild. ## Turn it on 1. **Select Deepgram for the ASR stage** In your `agent-config.yml`, set `asr_provider: deepgram` and select a model. If you do not set `asr_provider`, the cascade uses Deepgram anyway. Thus the main purpose of this step is to select the model. ```yaml agent-config.yml name: support-agent pipeline_mode: cascaded asr_provider: deepgram asr_model: nova-2 # forwarded onto the ASR socket llm_provider: openai llm_model: gpt-4o-mini tts_provider: elevenlabs tts_voice: language: en ``` 2. **Add your Deepgram key** Keys are **per-tenant**. Set `DEEPGRAM_API_KEY` in the FirstByte console at `agent.firstbyte.fast`. The control plane seals the key at rest and resolves it at call setup. The key is never stored in the config document. For single-tenant or local development, the runtime also reads the key from the environment: ```bash export DEEPGRAM_API_KEY=... ``` > **WARNING:** > If the ASR provider has no usable key, the runtime refuses the agent at session > creation with a clean error. The agent does not start and then fail mid-call. > Check that the key is present before you route live traffic. 3. **Place a test call** The control plane hydrates the config into the runtime with no rebuild. Place a call. Then open the [latency breakdown](/modalities/voice/observability/latency-breakdown) to see ASR-final timing for each turn. ## What the runtime sends Deepgram The FirstByte Agent Runtime opens one streaming Deepgram session for each call on the caller's shard. It forwards each 8 kHz PCM16 audio frame as the frame arrives off the media plane. The runtime consumes Deepgram's streaming results: interim hypotheses while the caller speaks, then a final transcript at the endpoint. The runtime sends the final text to the LLM stage. The runtime owns endpointing (see below). Thus it drives Deepgram in streaming mode and does not delegate turn-taking to the transcriber. The runtime passes the `asr_model` value directly to Deepgram. Any current Deepgram streaming model id works: ```yaml asr_provider: deepgram asr_model: nova-2 ``` Deepgram's general-purpose streaming model. This is the recommended default for English telephony. ```yaml asr_provider: deepgram asr_model: nova-3 ``` A newer model id. The runtime forwards it verbatim to Deepgram. Check accuracy on a staging number before you switch production traffic. ```yaml asr_provider: deepgram asr_model: nova-2-phonecall # or another Deepgram model id ``` You can select any streaming model that Deepgram exposes by its id. The runtime does not keep an allow-list. ## Barge-in Barge-in works with a Deepgram cascade. The runtime's **local, on-device VAD** handles endpointing and interruption, not the transcriber. Thus the caller can talk over the agent, and the agent stops. This behavior does not depend on your ASR choice. The runtime uses **hold-and-confirm** barge-in. A short backchannel ("mhm", "ok") does not cancel the agent. Sustained speech past the confirm window cancels the agent. Tune the behavior in the turn-detection block: ```yaml agent-config.yml turn_detection: type: local # on-device VAD owns endpointing + barge-in silence_threshold_ms: 500 min_speech_ms: 300 barge_confirm_ms: 300 # sustain speech past this to interrupt the agent response_min_speech_ms: 600 ``` > **NOTE:** > Backchannel suppression is **acoustic**. It uses utterance duration and the > hold-and-confirm window, not the words spoken. Lexical gating ("stop" vs "ok") > is not built. See [turn detection](/modalities/voice/concepts/turn-detection) > for the full set of controls. See > [barge-in debugging](/modalities/voice/observability/barge-in-debugging) > to examine what fired on a live call. ## Deepgram for text-to-speech Deepgram also ships as a **text-to-speech** provider (the Aura voices). To use Deepgram on both ends of the cascade, also set the TTS stage: ```yaml agent-config.yml tts_provider: deepgram tts_voice: aura-asteria-en ``` See [BYO ASR / LLM / TTS](/modalities/voice/runtime/byo-asr-llm-tts) for the full provider matrix across all three stages. ## Related - **[BYO ASR / LLM / TTS](/modalities/voice/runtime/byo-asr-llm-tts)** — Assemble the full cascade with every ASR, LLM, and TTS provider in one config file. - **[Turn detection](/modalities/voice/concepts/turn-detection)** — The endpointing and barge-in model that runs in front of Deepgram. - **[Runtime configuration](/modalities/voice/runtime/configuration)** — Every agent-config.yml field: providers, turn detection, budgets, codecs. - **[Latency breakdown](/modalities/voice/observability/latency-breakdown)** — See per-stage ASR / LLM / TTS timing on a live call.