Usage guide
Setup, environment variables, and Python/JavaScript/Go usage examples.
DeepgramSTT
Speech-to-text with Deepgram’s streaming Nova models. Good for low-latency English and multilingual transcription with keyword boosting and smart formatting. Authenticates with DEEPGRAM_API_KEY from your environment, or an explicitapi_key.
Fields
str
default:"UNSET"
Deepgram model to transcribe with, e.g. “nova-2” or “nova-3”. For Flux models use DeepgramSTTV2. Defaults to ‘nova-2’.
str
default:"UNSET"
BCP-47 language code to transcribe, e.g. “en-US” or “es”. Use “multi” with Nova-3 for code-switching speech. Defaults to ‘en-US’.
bool
default:"True"
Send partial transcripts to the client while the user is still speaking, in addition to final ones. Defaults to True.
bool
default:"UNSET"
Add punctuation and capitalization to transcripts. Defaults to True.
bool
default:"UNSET"
Format numbers, dates, currencies, emails and similar entities for readability. Defaults to True.
int | None
default:"UNSET"
Sample rate of incoming audio in Hz. Leave unset to detect it automatically. Defaults to None (not set).
int
default:"UNSET"
Milliseconds of silence before Deepgram finalizes an utterance. Set 0 to turn endpointing off. Defaults to 50.
bool
default:"UNSET"
Include filler words such as “uh” and “um” in transcripts. Defaults to True.
list[str] | None
default:"UNSET"
Words to boost or suppress, optionally with an intensifier, e.g. “kansas:-10”. Nova-2 and older models only; up to 100 entries. Defaults to None (not set).
list[str] | None
default:"UNSET"
Key terms or phrases to improve recognition of, e.g. “tretinoin”. Nova-3 only; up to 500 tokens in total. Defaults to None (not set).
bool
default:"UNSET"
Mask profanity in transcripts. Defaults to False.
bool
default:"UNSET"
Write spoken numbers as digits, e.g. “nine” becomes “9”. Defaults to False.
Union[str, List[str]] | None
default:"UNSET"
One or more labels attached to requests for usage reporting in the Deepgram console. Defaults to None (not set).
bool
default:"UNSET"
Detect speaker changes and label each word with a speaker. Defaults to False.
str
default:"UNSET"
Deepgram streaming endpoint URL. Change it only for a self-hosted or regional deployment. Defaults to ‘wss://api.deepgram.com/v1/listen’.
int | None
default:"UNSET"
Number of audio channels the agent receives. Leave unset to detect it automatically. Defaults to None (not set).
DeepgramSTTV2
Speech-to-text with Deepgram Flux, which combines transcription with built-in end-of-turn detection. Choose it for natural, low-latency turn-taking in voice agents. Authenticates with DEEPGRAM_API_KEY from your environment, or an explicitapi_key.
Fields
str
default:"UNSET"
Flux model to transcribe with: “flux-general-en” for English or “flux-general-multi” for multilingual speech. Defaults to ‘flux-general-en’.
int | None
default:"UNSET"
Sample rate of incoming audio in Hz. Leave unset to detect it automatically. Defaults to None (not set).
int | None
default:"UNSET"
Sample rate in Hz that audio is resampled to before it is sent to Deepgram, e.g. 16000. Defaults to None (not set).
float
default:"UNSET"
Confidence, 0 to 1, at which Flux signals an early end of turn. Lower values react sooner but more often prematurely. Defaults to 0.6.
float
default:"UNSET"
Confidence, 0 to 1, required before Flux declares the user’s turn finished. Higher values wait for more certainty. Defaults to 0.8.
int
default:"UNSET"
Milliseconds of silence after which the turn ends regardless of confidence. Defaults to 7000.
list[str] | None
default:"UNSET"
Key terms or phrases to improve recognition of, e.g. “tretinoin” or “customer service”. Up to 500 tokens in total. Defaults to None (not set).
list[str] | None
default:"UNSET"
Language codes to bias recognition toward, e.g. [“en”, “es”]. Only applies to flux-general-multi. Defaults to None (not set).
Union[str, List[str]] | None
default:"UNSET"
One or more labels attached to requests for usage reporting in the Deepgram console. Defaults to None (not set).
str
default:"UNSET"
Deepgram streaming endpoint URL. Change it only for a self-hosted or regional deployment. Defaults to ‘wss://api.deepgram.com/v2/listen’.
bool
default:"UNSET"
Start generating a reply on an early end-of-turn signal to cut response latency. The draft is discarded if the user keeps talking. Defaults to False.
int | None
default:"UNSET"
Number of audio channels the agent receives. Leave unset to detect it automatically. Defaults to None (not set).
bool
default:"True"
Send partial transcripts to the client as well as final ones. Turn detection and interruptions work either way. Defaults to True.