Assist Pipelines

The Assist pipeline integration runs the common steps of a voice assistant:

1. Wake word detection
2. Speech to text
3. Intent recognition
4. Text to speech

Pipelines are run via a WebSocket API:

{
  "type": "assist_pipeline/run",
  "start_stage": "stt",
  "end_stage": "tts",
  "input": {
    "sample_rate": 16000,
  }
}

The following input fields are available:

Name	Type	Description
start_stage	enum	Required. The first stage to run. One of wake_word, stt, intent, tts.
end_stage	enum	Required. The last stage to run. One of stt, intent, tts.
input	dict	Depends on start_stage: wake_word only: timeout - seconds before wake word detection times out (int, default: 3) noise_suppression_level - amount of noise suppression (int, 0 = disabled, 4 = max) auto_gain_dbfs - automatic gain (int, 0 = disabled, 31 = max) volume_multiplier - fixed volume amplification (float, 1.0 = no change, 2.0 = twice as loud) wake_word and stt: sample_rate - sample rate of incoming audio (int, hertz) intent and tts: text - input text (string)
pipeline	string	Optional. ID of the pipeline (use assist_pipeline/pipeline/list to get names).
conversation_id	string	Optional. Unique id for conversation.
timeout	number	Optional. Number of seconds before pipeline times out (default: 300).

Events

As the pipeline runs, it emits events back over the WebSocket connection. The following events can be emitted:

Name	Description	Emitted	Attributes
run-start	Start of pipeline run	always	pipeline - ID of the pipeline language - Language used for pipeline runner_data - Extra WebSocket data: stt_binary_handler_id is the prefix to send speech data over. timeout is the max run time for the whole pipeline.
run-end	End of pipeline run	always
wake_word-start	Start of wake word detection	audio only	engine: wake engine used metadata: incoming audio timeout: seconds before wake word timeout metadata
wake_word-end	End of wake word detection	audio only	wake_word_output - Detection result data: wake_word_id is the id of detected wake word timestamp is the detection time relative to start of audio stream (milliseconds, optional)
stt-start	Start of speech to text	audio only	engine: STT engine used metadata: incoming audio metadata
stt-vad-start	Start of voice command	audio only	timestamp: time relative to start of audio stream (milliseconds)
stt-vad-end	End of voice command	audio only	timestamp: time relative to start of audio stream (milliseconds)
stt-end	End of speech to text	audio only	stt_output - Object with text, the detected text.
intent-start	Start of intent recognition	always	engine - Agent engine used language: Processing language. intent_input - Input text to agent
intent-end	End of intent recognition	always	intent_output - conversation response
tts-start	Start of text to speech	audio only	engine - TTS engine used language: Output language. voice: Output voice. tts_input: Text to speak.
tts-end	End of text to speech	audio only	media_id - Media Source ID of the generated audio url - URL to the generated audio mime_type - MIME type of the generated audio
error	Error in pipeline	on error	code - Error code (see below) message - Error message

Error codes

The following codes are returned from the pipeline error event:

• wake-engine-missing - No wake word engine is installed
• wake-provider-missing - Configured wake word provider is not available
• wake-stream-failed - Unexpected error during wake word detection
• wake-word-timeout - Wake word was not detected within timeout
• stt-provider-missing - Configured speech-to-text provider is not available
• stt-provider-unsupported-metadata - Speech-to-text provider does not support audio format (sample rate, etc.)
• stt-stream-failed - Unexpected error during speech-to-text
• stt-no-text-recognized - Speech-to-text did not return a transcript
• intent-not-supported - Configured conversation agent is not available
• intent-failed - Unexpected error during intent recognition
• tts-not-supported - Configured text-to-speech provider is not available or options are not supported
• tts-failed - Unexpected error during text-to-speech

Sending speech data

After starting a pipeline with stt as the first stage of the run and receiving a stt-start event, speech data can be sent over the WebSocket connection as binary data. Audio should be sent as soon as it is available, with each chunk prefixed with a byte for the stt_binary_handler_id.

For example, if stt_binary_handler_id is 1 and the audio chunk is a1b2c3, the message would be (in hex):

stt_binary_handler_id
||
01a1b2c3
  ||||||
   audio

To indicate the end of sending speech data, send a binary message containing a single byte with the stt_binary_handler_id.

Wake word detection

When start_stage is set to wake_word, the pipeline will not run until a wake word has been detected. Clients should avoid unnecessary audio streaming by using a local voice activity detector (VAD) to only start streaming when human speech is detected.

For wake_word, the input object should contain a timeout float value. This is the number of seconds of silence before the pipeline will time out during wake word detection (error code wake-word-timeout). If enough speech is detected by Home Assistant's internal VAD, the timeout will be continually reset.

Audio Enhancements

The following settings are available as part of the input object when start_stage is set to wake_word:

• noise_suppression_level - level of noise suppression (0 = disabled, 4 = max)
• auto_gain_dbfs - automatic gain control (0 = disabled, 31 = max)
• volume_multiplier - audio samples multiplied by constant (1.0 = no change, 2.0 = twice as loud)

If your device's microphone is fairly quiet, the recommended settings are:

• noise_suppression_level - 2
• auto_gain_dbfs - 31
• volume_multiplier - 2.0

Increasing noise_suppression_level or volume_multiplier may cause audio distortion.