Recognizer

Parameters

Speech vendor to use (see list below). Note: this field is case sensitve, all the built-in vendors are lower case eg aws not AWS

(Google, Microsoft) An array of alternative languages that the speaker may be using.

DTMF key that terminates continuous ASR.

Timeout value for continuous ASR.

(AssemblyAI) When true, automatically sets assemblyAiOptions.prompt to the text of the nested say Function of the parent gather. Only applies when there is a nested say and no explicit assemblyAiOptions.prompt has been supplied. SSML tags are stripped before use.

Custom service endpoint to connect to instead of hosted Microsoft regional endpoints.

(Google) Enable speaker diarization.

(Google) Set the maximum speaker count.

(Google) Set the minimum speaker count.

(Google) Use an enhanced model.

(AWS) The method to use when filtering speech: remove, mask, or tag.

(Google, Microsoft, Deepgram, Nvidia, Soniox) Array of words or phrases to assist speech detection.
See Providing speech hints below.

(Google, Nvidia) Number indicating the strength to assign to the configured hints.
See examples below.

(AWS) Enable channel identification.

(Microsoft) Initial speech timeout in milliseconds.

(Google) Set the interaction type: discussion, presentation, phone_call, voicemail,
professionally_produced, voice_search, voice_command, dictation.

If true, interim transcriptions are sent.
Default: false.

Note: this only effects use in a Transcribe Function, in Gather interims are sent based on the presence of a partialResponseHook

Language code to use for speech detection.
Defaults to the application-level setting.

(AWS) The name of the custom language model when processing speech.

If provided, final transcripts with confidence lower than this value
return a reason of 'stt-low-confidence' in the webhook.

(Google) Speech recognition model to use.
Default: phone_call.

(Google) Set an industry NAICS code that is relevant to the speech.

(Microsoft) simple or detailed.
Default: simple.

(Google, Deepgram, Nuance, Nvidia) If true, filter profanity from speech transcription.
Default: false.

(Microsoft) masked, removed, or raw.
Default: raw.

(Google) Enable automatic punctuation.

(Microsoft) Request signal-to-noise ratio information.

If true, recognize both caller and called party speech using separate recognition sessions.

(Google) If true, return only a single utterance/transcript.
Default: true for gather.

Webhook to receive an HTTP POST when an interim or final transcription is received.

If true, delay connecting to the cloud recognizer until speech is detected.

If vad is enabled, this setting governs the sensitivity of the voice activity detector;
value must be between 0 and 3 inclusive.
Lower numbers mean more sensitivity.

If vad is enabled, the number of milliseconds of speech required before connecting to the cloud recognizer.

(AWS) The name of a vocabulary filter to use when processing the speech.

(AWS) The name of a vocabulary to use when processing the speech.

Vendor-specific options

assemblyAiOptions

AssemblyAI API key to authenticate with (overrides the credential configured for your account).

The version of the streaming API to use, v2 or v3.

AssemblyAI speech model to use for recognition (e.g., u3-rt-pro).

Format each speaker turn. Default: true.

Confidence threshold (0.0–1.0) the model must meet before triggering an end-of-turn. Example: 0.5.

Minimum silence duration (milliseconds) required after the last spoken word before the system triggers an end-of-turn event, when the model is confident an EOT should occur. Default: 400. Example: 500.

Maximum silence duration (milliseconds) allowed in a turn before an end-of-turn is automatically triggered. Default: 1280. Example: 2000.

Minimum silence duration (milliseconds) before allowing end-of-turn.

List of key terms to boost in recognition. Example: ["Wipple", "Miraicom"].

Prompt text to guide the recognition model — typically a description of the expected user response. Improves accuracy when the response is predictable (e.g., the text of the question you just asked). See also autogeneratePrompt on the recognizer, which auto-populates this from a nested say in a gather Function.

Enable automatic language detection.

Voice activity detection threshold.

Timeout (seconds) for inactivity before closing the stream.

azureOptions

Duration (in milliseconds) of non-speech audio within a phrase that's currently being spoken before that phrase is considered "done."
See here for details.

Enable audio logging. See here for details

deepgramOptions

Number of alternative transcripts to return.

Deepgram API key to authenticate with (overrides the credential configured for your account).

ID of custom model.

Whether to assign a speaker to each word in the transcript.

If set to '2021-07-14.0', the legacy diarization feature will be used.

Indicates the number of milliseconds of silence Deepgram
will use to determine a speaker has finished saying a word or phrase.
Value must be either a number of milliseconds or 'false' to disable the feature entirely.
Default: 10ms.

Available with Deepgram Flux only. The confidence threshold for triggering EagerEndOfTurn events, enableing early LLM response generation.

Available with Deepgram Flux only. The confidence threshold for end-of-turn detection

Available with Deepgram Flux only. Maximum silence duration before forcing an EndOfTurn, regardless of confidence. Valid values are between 500 and 10000 milliseconds.

An array of keyterm prompts

An array of keywords
to which the model should pay particular attention to boosting or suppressing to help it understand context.

Indicates whether to opt out of Deepgram's model improvement program.

Deepgram model used to process submitted audio.
Example models: 'nova-3', 'nova-2', 'nova-2-phonecall'; see Deepgram docs for full list.
Default: 'general'.

Indicates whether to transcribe each audio channel independently.

Indicates whether to enable Deepgram's no_delay feature.

Indicates whether to convert numbers
from written format (e.g., "one") to numerical format (e.g., "1").

Indicates whether to remove profanity
from the transcript.

Indicates whether to add punctuation
and capitalization to the transcript.

Whether to redact information
from transcripts.
Allowed values: 'pci', 'numbers', 'true', 'ssn'.

An array of terms or phrases
to search for in the submitted audio and replace.

An array of terms or phrases to search for in the submitted audio.

Causes a transcript to be returned as soon as Deepgram's is_final property is set.
This should only be used in scenarios where you expect a very short confirmation
or directed command and want minimal latency.

Indicates whether to enableDeepgram's Smart Formatting feature.

A tag to associate with the request.
Tags appear in usage reports.

Deepgram tier you would like to use.
Allowed values: 'enhanced', 'base'.
Default: 'base'.

A number of milliseconds of silence that Deepgram will wait
after the last word was spoken before returning an UtteranceEnd event,
which is used by Wipple CPaaS to trigger the transcript webhook if this property is supplied.
This is essentially Deepgram's version of continuous ASR.

Deepgram version of the model to use.
Default: 'latest'.

googleOptions

Google serves three speech APIs from one set of credentials: Cloud Speech-to-Text v1, Cloud Speech-to-Text v2, and the Gemini Live transcription models. Which one runs is chosen by the recognizer model: a name beginning with gemini- selects the Gemini Live API, anything else a Cloud Speech recognizer. Leave model unset and Wipple CPaaS picks a Cloud Speech model appropriate to the language, as it always has. A default model can also be stored on the Google speech credential configured for your account, which applies to every call on that account unless the request overrides it.

Gemini Live transcription does not support speaker diarization or word timestamps, and Google limits a live session to 10 minutes; the diarization and words recognizer properties are ignored on that path.

Recognition runs on the GA Vertex AI interface, authenticating as the speech credential's own service account: no additional secret is needed, but the project must have billing enabled and the service account roles/aiplatform.user. The project comes from the service account and the region from parentPath, defaulting to global, where these models are served. Name the model as Vertex AI publishes it — the preview Gemini Developer API ("AI Studio") spells the same models differently and is not supported, since it rejects service accounts.

Gemini transcription style, for the gemini models only: VERBATIM (default) transcribes speech as spoken, SMART cleans up disfluencies and applies formatting.

Up to 1000 terms, acronyms, brand names or proper nouns to bias recognition towards, for the gemini models only. Example: ["Wipple", "Miraicom"]. When omitted, the recognizer's hints are used instead.

Cloud Speech-to-Text API version, v1 (default) or v2. Ignored by the gemini models, which are a different API.

ID of a Google Speech recognizer resource (v2 only).

Parent resource path for the Google Speech recognizer (v2 only), e.g. projects/my-project/locations/us-central1. A location other than global also routes the request to that region's endpoint.

Milliseconds to wait for speech to start (v2 only).

Milliseconds of silence used to detect the end of speech (v2 only).

Enable Google's voice activity detection events (v2 only).

Transcript normalization rules (v2 only), each {search, replace, caseSensitive}.

xaiOptions

xAI (xai) is a live streaming recognizer: both interim and final transcripts are supported.

xAI API key to authenticate with (overrides the credential configured in Wipple CPaaS).

Transcription model to use, e.g. grok-voice-transcribe-2.0 or grok-voice-transcribe-1.0.
If not supplied, xAI selects its own current default model.

Number of milliseconds of silence before an utterance is considered final.
Valid values are between 0 and 5000.

Enable speaker diarization.

Include filler words (e.g. "um", "uh") in the transcript.

An array of key terms to boost recognition.

End-of-turn confidence threshold, between 0.0 and 1.0.

Maximum number of milliseconds of silence before forcing an end-of-turn final result.
Valid values are between 1 and 5000.

houndifyOptions

Latitude of caller.

Longitude of caller.

City of caller.

State of caller.

Country of caller.

Timezone of caller.

ibmOptions

ID of a custom acoustic model.

Base model to be used.

IBM speech instance ID (overrides the credential configured for your account).

ID of a custom language model.

The model to use for speech recognition.

IBM API key to authenticate with (overrides the credential configured for your account).

IBM region (overrides the credential configured for your account).

Set to true to prevent IBM from using your API request data to improve their service.

A tag value
to apply to the request data provided.

nuanceOptions

When true, custom resources (DLMs, wordsets, etc.) can use the entire weight range.

An object containing arbitrary key-value pairs to inject into the call log.

Nuance client ID to authenticate with (overrides the credential configured for your account).

If speaker profiles are used, whether to discard updated speaker data.
By default, data is stored.

Whether to remove the wakeup word from the final result.

Object containing key-value pairs of formatting options and values defined in the data pack.

Keyword for a formatting type defined in the data pack.

Whether to include a tokenized recognition result.

Endpoint of the on-prem Krypton endpoint to connect to.
Default: Hosted service.

Whether to terminate recognition when failing to load external resources.

Maximum number of n-best hypotheses to return.

Maximum silence (in milliseconds) allowed while waiting for user input after recognition timers are started.

Whether to enable auto-punctuation.

Maximum duration (in milliseconds) of the recognition turn.

An array of zero or more recognition resources
(domain LMs, wordsets, etc.) to improve recognition.

Name of a built-in resource in the data pack.

An external DLM or settings file
for creating or updating a speaker profile.

An object containing HTTP cache-control directives (e.g., max-age, etc.).

When true, allow transcription to proceed even if resource loading fails.

Time to wait when downloading resources.

Resource type: 'undefined_resource_type', 'wordset', 'compiled_wordset', 'domain_lm',
'speaker_profile', 'grammar', 'settings'.

Location of the resource as a URN reference.

Inline grammar in SRGS XML format.

Inline wordset JSON resource.
See Wordsets for details.

Whether the resource will be used multiple times.
Allowed values: 'undefined_reuse', 'low_reuse', 'high_reuse'.
Default: low_reuse.

Input field setting the weight of the
domain LM or built-in resource relative to the data pack.
Allowed values: 'defaultWeight', 'lowest', 'low', 'medium', 'high', 'highest'.
Default: MEDIUM.

Weight of the DLM or built-in resource as a numeric value from 0 to 1.
Default: 0.25.

Array of wakeup words.

The level of recognition results: 'final', 'partial', 'immutable_partial'.
Default: final.

Nuance secret to authenticate with (overrides the credential configured for your account).

A balance between detecting speech and noise (breathing, etc.), ranging from 0 to 1.
0 means ignore all noise, 1 means interpret all noise as speech.
Default: 0.5.

Mapping to internal weight sets for language models in the data pack.

Whether to disable call logging and audio capture.
By default, call logs, audio, and metadata are collected.

When true, the first word in a sentence is not automatically capitalized.

Specialized language model.

How many sentences (utterances) within the audio stream are processed.
Allowed values: 'single', 'multiple', 'disabled'.
Default: single.

Minimum silence (in milliseconds) that determines the end of a sentence.

Identifies a specific user within the application.

nvidiaOptions

An object of key-value pairs that can be sent to Nvidia for custom configuration.

Number of alternative transcripts to return.

Indicates whether to remove profanity from the transcript.

Indicates whether to provide punctuation in the transcripts.

GRPC endpoint (ip:port) that Nvidia Riva is listening on.

Indicates whether to provide verbatim transcripts.

Indicates whether to provide word-level detail.

openaiOptions

OpenAI API key to authenticate with (if provided, this overrides the credential configured for your account).

OpenAI language code. Language could also be set at the recognizer level, but if set here this value will take precedence. Note that language code is supplied in ISO-639-1 (e.g. en) format.

OpenAI model. Currently 'whisper-1', 'gpt-4o-mini-transcribe', 'gpt-4o-transcribe', 'gpt-realtime-whisper', and 'gpt-live-transcribe' are supported. Note: 'gpt-live-transcribe' and 'gpt-realtime-whisper' do not support OpenAI's server-side turn detection. Any turn_detection settings are ignored for these models; Wipple CPaaS detects end of speech locally instead (see vadMode, vadSilenceMs, vadVoiceMs).

Used only with 'gpt-live-transcribe', which accepts a list of language hints rather than a single code, e.g. ["en", "fr"]. Takes precedence over language and over the recognizer-level language.

Used only with 'gpt-live-transcribe'. Literal terms that may appear in the audio — product names, acronyms, medications. These are hints, not required output. If you do not set a value here but you do provide hints, the hints are sent as keywords.

Used only with 'gpt-live-transcribe'. One of 'minimal', 'low', 'medium', 'high', or 'xhigh' — the latency/accuracy trade-off. Lower values emit partial text sooner; higher values give the model more audio context and can improve accuracy.

An optional text to guide the model's style or continue a previous audio segment. Note that if you do not set a value here but you do provide hints, a prompt will automatically be generated to convey the hints to the OpenAI model.

Governs whether noise reduction is applied to the input audio.

'near_field' or 'far_field'.

Templates that can be used to dynamically construct the prompt to be sent to OpenAI.

A template that be be used to construct a prompt to convey hints. Use the placeholder "{{hints}}", e.g. "Please transcribe the following audio, making sure to spell the following words correctly: {{hints}}"

A template that can be used to construct a prompt based on the recent conversation. Use the placeholder "{{turns}}", e.g. "Here is the recent conversation history: {{turns}}".
By default, the last 4 turns of the conversation will be used, but you can change this as follows: "Here is the recent conversation history: {{turns:3}}".

Specifies how to detect when a speaker has finished speaking.

'none', 'server_vad', or 'semantic_vad'.

'low', 'medium', 'high, or 'auto'. Used only for semantic_vad mode. The eagerness of the model to respond. low will wait longer for the user to continue speaking, high will respond more quickly. auto is the default and is equivalent to medium.

Used only for server_vad mode. Activation threshold for VAD (0.0 to 1.0), this defaults to 0.5. A higher threshold will require louder audio to activate the model, and thus might perform better in noisy environments.

Used only for server_vad mode. Amount of audio to include before the VAD detected speech (in milliseconds). Defaults to 300ms.

Used only for server_vad mode. Duration of silence to detect speech stop (in milliseconds). Defaults to 500ms. With shorter values the model will respond more quickly, but may jump in on short pauses from the user.

Used only with 'gpt-live-transcribe' or 'gpt-realtime-whisper'. Sensitivity of the local voice activity detector; integer 0-3. Higher is more aggressive. Defaults to 2.

Used only with 'gpt-live-transcribe' or 'gpt-realtime-whisper'. Milliseconds of silence required to detect end-of-speech and commit the buffer. Defaults to 500ms.

Used only with 'gpt-live-transcribe' or 'gpt-realtime-whisper'. Milliseconds of voice required to detect start-of-speech. Defaults to 250ms.

houndifyOptions

Latitude coordinate for location-based queries.

Longitude coordinate for location-based queries.

City name for location context.

State or province for location context.

Country for location context.

Time zone identifier (e.g., 'America/New_York').

Houndify domain to enable specific capabilities.

Custom audio streaming endpoint URL.

Maximum silence duration (in seconds) before ending the request.

Maximum silence duration (in seconds) after detecting a complete query.

Maximum silence duration (in seconds) after detecting a partial query.

Voice activity detection sensitivity level.

Voice activity detection timeout value.

Voice activity detection mode.

Minimum voice duration (in milliseconds) for VAD detection.

Minimum silence duration (in milliseconds) for VAD detection.

Enable VAD debugging information.

Audio format specification for the audio stream.

Enable audio noise reduction processing.

Filter profanity from transcription results.

Enable automatic punctuation in transcripts.

Enable automatic capitalization in transcripts.

Minimum confidence threshold for accepting recognition results.

Filter disfluencies (um, uh, etc.) from transcription results.

Maximum number of results to return.

Include word-level timestamps in transcription results.

Maximum number of alternative transcriptions to return.

Interval (in milliseconds) for sending partial transcript updates.

Session timeout duration (in seconds).

Connection timeout duration (in seconds).

Array of custom words or phrases to improve recognition accuracy.

Custom language model identifier to use for recognition.

sonioxOptions

Soniox API key.

Soniox model to use.
Default: 'precision_ivr'.

Indicates whether to remove profanity from the transcript.

Properties that dictate whether to store audio and/or transcripts.
Can be useful for debugging purposes.

If true, do not allow search.

If true, do not store audio.

If true, do not store transcripts.

Storage identifier.

Storage title.

gladiaOptions

The processing model to use. Default: 'solaria-1'.

Duration of silence (in seconds) that triggers utterance completion. Value must be between 0.01 and 10. Default: 0.05.

Maximum duration (in seconds) without endpointing detection. Value must be between 5 and 60. Default: 5.

Attach custom metadata to the transcription session.

Language configuration for the transcription.

Array of language codes to use for transcription.

Enable automatic language switching within the audio stream.

Audio pre-processing configuration.

Enable audio enhancement processing.

Speech detection threshold value.

Real-time processing configuration.

Array of custom vocabulary terms to improve recognition accuracy.

Configuration for custom vocabulary behavior.

Array of custom spelling rules.

Configuration for custom spelling behavior.

Enable real-time translation.

Translation configuration including target languages.

Enable named entity recognition (NER).

Enable sentiment analysis on transcriptions.

Post-processing configuration.

Enable automatic summarization.

Configuration for summarization behavior.

Enable automatic chapter detection.

Control which event types are received during the session.

Receive pre-final transcription events.

Receive final transcription events.

Receive speech recognition events.

Enable callback delivery for transcription events. Default: false.

Callback configuration when callback is enabled.

URL to receive callback events.

speechmaticsOptions

Audio events to report.

"applause", "laughter", or "music"

Audio transcription configuration.

Additional vocabulary words.

Audio filtering configuration.

Enable partial transcriptions.

Language to transcribe.

"fixed" or "flexible"

Deprecated by Speechmatics in favor of model. "standard" or "enhanced".

Model selection (replaces the deprecated operating_point), e.g. "standard", "enhanced", or "melia-1". See Speechmatics model docs for the current list.

Punctuation configuration

Transcript filtering configuration. See Speechmatics formatting docs for details.

Remove hesitation sounds ("um", "uh") from the transcript. English only. Default: false.

Word replacement rules applied to the transcript. Each entry is an object with from (a plain word or a /regex/) and to (the replacement text). Example: [{"from": "foo", "to": "bar"}].

Configuration for conversation-based transcription features.

Duration of silence (in seconds) that triggers an end-of-utterance event. This controls how long the system waits after the speaker stops talking before determining that the utterance is complete. See Speechmatics turn detection docs for details.

Audio filtering configuration.

Volume threshold to filter.

Providing speech hints

Many recognizers support the ability to provide a dynamic list of words or phrases that should be "boosted" by the recognizer, i.e. the recognizer should be more likely to detect this terms and return them in the transcript. A boost factor can also be applied. In the most basic implementation it would look like this:

"hints": ["benign", "malignant", "biopsy"],
"hintsBoost": 50

Additionally, google and nvidia allow a boost factor to be specified at the phrase level, e.g.

"hints": [
  {"phrase": "benign", "boost": 50},
  {"phrase": "malignant", "boost": 10},
  {"phrase": "biopsy", "boost": 20},
]