Usage guide
Setup, environment variables, and Python/JavaScript/Go usage examples.
AssemblyAISTT
Speech-to-text with AssemblyAI’s streaming Universal models, using your own AssemblyAI key. Good for accurate real-time transcription with built-in end-of-turn detection, in English or multilingual. Authenticates with ASSEMBLYAI_API_KEY from your environment, or an explicitapi_key.
Fields
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 to send to AssemblyAI, 8000 to 96000. Leave unset to send audio at the rate it arrives, with no resampling. Defaults to None (not set).
Literal['pcm_s16le']
default:"UNSET"
Encoding of the audio sent to AssemblyAI. Only “pcm_s16le” (16-bit little-endian PCM) is supported. Defaults to ‘pcm_s16le’.
bool
default:"UNSET"
Whether finished turns come back with punctuation and casing applied. Universal-3 Pro models always format turns and ignore this. Defaults to True.
list[str] | None
default:"UNSET"
Words or phrases to boost recognition of, such as product names, jargon or proper nouns. Defaults to None (not set).
float | None
default:"UNSET"
Confidence, 0 to 1, needed to end a turn; higher waits for more certainty. Universal-streaming models only (0.5 used when unset); Pro models ignore it. Defaults to None (not set).
int | None
default:"UNSET"
Silence in milliseconds before a confident end of turn is finalized. When unset, universal-streaming models use 800. Defaults to None (not set).
int | None
default:"UNSET"
Silence in milliseconds after which a turn ends regardless of confidence. When unset, universal-streaming models use 2000. Defaults to None (not set).
Literal['universal-streaming-english', 'universal-streaming-multilingual', 'universal-3-5-pro', 'universal-3-6-pro', 'u3-rt-pro', 'u3-rt-pro-beta-1']
default:"UNSET"
Recognition model: “universal-streaming-english”, “universal-streaming-multilingual”, or Universal-3 Pro “universal-3-5-pro” / “universal-3-6-pro”. (“u3-rt-pro” is a retired alias). Defaults to ‘universal-streaming-english’.
bool | None
default:"False"
Whether to detect the spoken language automatically. Supported by the multilingual model. Defaults to False.
str | list[str] | None
default:"UNSET"
Expected language or languages, e.g. “es” or [“en”, “es”]. Up to 10 ISO 639-1 codes; “multi” cannot be combined with others. Defaults to None (not set).
Literal['min_latency', 'balanced', 'max_accuracy'] | None
default:"UNSET"
Accuracy versus latency preset: “min_latency”, “balanced” or “max_accuracy”. Universal-3 Pro models only. Defaults to None (not set).
str | None
default:"UNSET"
Free-text description of the audio, such as topic or vocabulary, used to steer transcription. Universal-3 Pro models only. Defaults to None (not set).
str | None
default:"UNSET"
What the agent just said, used to improve transcription of the user’s reply. Universal-3 Pro models only, up to 1750 characters. Defaults to None (not set).
int | None
default:"UNSET"
How many earlier conversation turns are carried forward as context for transcription. 0 turns this off. Defaults to None (not set).
bool | None
default:"UNSET"
Whether to send extra partial transcripts at a steady pace during long turns. Defaults to None (not set).
int | None
default:"UNSET"
How soon the first early partial transcript arrives, in milliseconds. Lower reacts faster to barge-in; higher is more reliable. Defaults to None (not set).
float | None
default:"UNSET"
Voice activity sensitivity, 0 to 1. Lower values pick up quieter speech. Defaults to None (not set).
Literal['near-field', 'far-field'] | None
default:"UNSET"
Isolate the main speaker and suppress background noise. Use “near-field” for headsets and phones, “far-field” for rooms and laptop mics. Defaults to None (not set).
float | None
default:"UNSET"
How aggressively background audio is suppressed, 0 to 1. Requires voice_focus. Defaults to None (not set).
bool
default:"UNSET"
Start generating a reply from early transcripts before the turn ends, to cut response latency. Only helps on universal-streaming models. Defaults to False.
str
default:"UNSET"
Data region for the AssemblyAI endpoint, “US” or “EU”. Ignored when base_url is set. Defaults to ‘US’.
str | None
default:"UNSET"
Custom streaming endpoint URL, e.g. for a private deployment. Overrides region. Defaults to None (not set).
int | None
default:"UNSET"
Deprecated; use min_turn_silence. If set, it takes precedence over min_turn_silence. Defaults to None (not set).
bool
default:"True"
Whether partial transcripts are also sent to the client alongside final ones. Turn detection uses partials either way. Defaults to True.