Deepgram speech provider

A Deepgram speech provider lets Live Hub use your Deepgram account for speech-to-text (STT), text-to-speech (TTS), or both.

Live Hub reaches Deepgram at api.deepgram.com. For text-to-speech, you can point it at your own private deployment instead by supplying that deployment's URL.

Deepgram's advanced configuration parameters are set on a bot connection rather than on the provider. See Configure advanced parameters at the bot connection level.

Add a Deepgram speech provider

To add a Deepgram speech provider:

1. In the 'Authentication key' field, enter the API key from your Deepgram account.

2. Optional: For text-to-speech, under Advanced, in the 'Text to Speech (TTS) URL' field, enter the URL of your private deployment.

   

3. Click Create.

Configure advanced parameters at the bot connection level

Deepgram's advanced configuration parameters go under sttPassthruConfig and ttsPassthruConfig in the JSON editor on the bot connection's Advanced tab, and Live Hub merges them into the request it sends to Deepgram. See Manage bot connections.

Speech-to-text

The full set of speech-to-text parameters is the query parameters of Deepgram's streaming speech-to-text API.

{
    "sttPassthruConfig": {
        "numerals": true
    }
}

For keyword boosting, use the sttSpeechContexts bot parameter rather than sttPassthruConfig.

Flux models use Deepgram's Flux API instead, which takes its own set of query parameters. None of the exclusions above apply to it. For the parameters that end a Flux turn, see Turn-taking.

Text-to-speech

The full set of text-to-speech parameters is the query parameters of Deepgram's text-to-speech API.

{
    "ttsPassthruConfig": {
        "tag": "example-tag"
    }
}

Turn-taking

Live Hub finds the end of a turn from the punctuation in Deepgram's recognitions. Deepgram returns final recognitions while the caller is still speaking, rather than holding them until the end of an utterance, and punctuates what it returns. A recognition ending in a period, a question mark, or an exclamation mark is treated as a finished thought, and one that does not end in any of them is given longer, in case the caller is still talking.

Two fields on the bot connection's Settings tab, under Speech-to-Text (STT) Settings, set how long Live Hub waits in each case. Each one has a bot parameter behind it, which you can set on the Advanced tab instead:

Field	Bot parameter
'End-of-turn timeout for recognitions without terminating punctuation (ms)'	continuousASRTimeoutInMS
'End-of-turn timeout for recognitions with terminating punctuation (ms)'	continuousASRPunctuationTimeoutMS

For their defaults and ranges, see Deepgram speech-to-text.

Both are worth tuning against your own callers. The wait after punctuation can be short, because the model has already signalled a complete thought, but too short a value cuts people off when they pause between sentences. The wait without punctuation is the safety net for speech the model did not resolve into a sentence, and too short a value ends the turn while the caller is still thinking.

Live Hub sets two further parameters to make this work. Override either on the Advanced tab if you need to:

Parameter	Value	What it does
continuousASR	true	Collects the recognitions into one turn instead of sending each of them to the bot.
endpointingMs	10	Closes each recognition almost at once. It governs the individual recognitions, not the end of the turn.

Flux models

Flux decides for itself when a turn has ended and tells Live Hub, so none of the fields and parameters above apply to it. Its own parameters go under sttPassthruConfig on the Advanced tab:

Parameter	What it does
eot_threshold	Sets how confident Flux must be that the turn is over before it says so, from 0.5 to 0.9. Default 0.7. Raise it to let callers pause for longer, lower it to end turns sooner.
eot_timeout_ms	Sets how long Flux waits before it ends the turn whatever its confidence. Default 5000 ms.
eager_eot_threshold	Signals a likely end of turn before Flux is fully confident, from 0.3 to 0.9. It shortens the pause before the bot replies, at the cost of 50% to 70% more requests to the bot.

{
    "sttPassthruConfig": {
        "eot_threshold": 0.8
    }
}

Training data

Selecting Deepgram adds a toggle to the bot connection's Settings tab: 'Allow Deepgram to use the audio data to improve its models'. Turning it on lets Deepgram use the call audio for model development, and Deepgram handles that data under its usage terms. Turning it off raises your usage costs.

Language, model, and voice

You select the language, the model, and the voice per bot connection, on its Settings tab, not on the provider. See Manage bot connections.

If the speech-to-text model supports it, a 'Multilingual' check box appears on the same tab. Selecting it changes the 'Language' field to 'Main language'.