# Deepgram

> Fast, accurate English and Indian English. Every listening setting for Deepgram, and how to set it.

Source: https://voxio.in/docs/voice-listening-deepgram · voxioagent 2.0.0 · Index: https://voxio.in/docs/llms.txt

Service: `deepgram-streaming`. A new setup starts from:

```json
{
  "service": "deepgram-streaming",
  "model": "nova-3",
  "language": "en-IN"
}
```

## Set it

On the flow's **Voice** tab, choose **Listening**, then **Deepgram**. Or from code:

```js
await voxio.flows.update(flowKey, {
  stt: {"service":"deepgram-streaming","model":"nova-3","language":"en-IN"},
});
```

## Settings

Settings marked “under More settings” are behind **More settings** in the dashboard.

- `model` (string): **Model**. Which speech model listens. Hover an option to compare. Default: `"nova-3"`.
- `language` (string): **Language**. The language callers speak. Default: `"en-IN"`. Only applies to some models.
- `language-hint` (string): **Language hint** (under More settings). The language callers most often use. Only applies to some models.
- `keyterm` (string[]): **Boost words**. Words the agent should never mishear.
- `endpointing` (number (ms)): **End-of-speech silence** (under More settings). Silence before a sentence is considered finished. From 10 to 5000. Default: `300`. Only applies to some models.
- `utterance-end-ms` (number (ms)): **Utterance end** (under More settings). Longest gap allowed inside one sentence. From 1000 to 1000. Default: `1000`. Only applies to some models.
- `eot-threshold` (number): **Turn-end confidence** (under More settings). How sure the model must be that the caller finished. From 0.5 to 1. Default: `0.7`. Only applies to some models.
- `eager-eot-threshold` (number): **Early reply confidence** (under More settings). Lets the agent start answering early. From 0.3 to 0.9. Only applies to some models.
- `eot-timeout-ms` (number (ms)): **Turn timeout** (under More settings). Longest the agent waits before replying anyway. From 500 to 60000. Default: `5000`. Only applies to some models.
- `smart-format` (boolean): **Smart formatting** (under More settings). Numbers, dates and currency as digits. Default: `false`. Only applies to some models.
- `punctuate` (boolean): **Punctuation** (under More settings). Add commas, full stops and question marks. Default: `false`. Only applies to some models.
- `numerals` (boolean): **Numbers as digits** (under More settings). Write “twenty five” as 25. Default: `false`. Only applies to some models.
- `interim-results` (boolean): **Live partial results** (under More settings). Transcribe while the caller is still talking. Default: `true`. Only applies to some models.
- `vad-events` (boolean): **Speech-start events** (under More settings). Signal the moment the caller starts speaking. Default: `false`. Only applies to some models.
- `profanity-filter` (boolean): **Filter profanity** (under More settings). Mask swear words in transcripts. Default: `false`. Only applies to some models.
- `diarize` (boolean): **Tell speakers apart** (under More settings). Label who said what when more than one person talks. Default: `false`. Only applies to some models.
- `redact` (string[]): **Redact** (under More settings). Hide card numbers, SSNs or all numbers from transcripts. Only applies to some models.

## Choices

### Model choices

| Value | Name | Notes |
| --- | --- | --- |
| `nova-3` | Nova 3 | Best general accuracy. The safe default. Accurate on English and Indian English, fast enough for live calls. |
| `nova-3-medical` | Nova 3 Medical | Knows drug names, conditions and procedures. Use for clinics and pharmacies. |
| `nova-2` | Nova 2 | Previous generation. Only pick it if Nova 3 struggles with a specific accent. |
| `nova-2-phonecall` | Nova 2 Phone call | Tuned for 8 kHz calls. Trained on low-quality phone audio. Can help on noisy mobile lines. |
| `flux-general-en` | Flux (English) | Built-in turn detection. Knows when the caller has finished speaking, so replies feel snappier. English only. |
| `flux-general-multi` | Flux (Multilingual) | Built-in turn detection. Flux turn detection for callers who switch languages. |

### Language choices (for some models)

| Value | Name | Notes |
| --- | --- | --- |
| `en-IN` | English (India) |  |
| `en` | English |  |
| `en-US` | English (US) |  |
| `en-GB` | English (UK) |  |
| `hi` | Hindi |  |
| `multi` | Multilingual | Code-switching |

## Next steps

- [Listening](https://voxio.in/docs/voice-listening.md): compare the other providers.
- [Turn-taking](https://voxio.in/docs/turn-taking.md): when the agent answers.