Skip to content

Voice Control

Jump to bottom Edit New page
knoop7 edited this page Jun 22, 2026 · 5 revisions

Voice Control

Voice control is the core feature of Ava, allowing you to control smart home devices by speaking.

How It Works

┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  You speak  │ -> │ Ava records │ -> │Home Assistant│ -> │ Ava plays   │
│ wake word + │    │ sends audio │    │   speech    │    │   voice     │
│   command   │    │             │    │ recognition │    │   reply     │
└─────────────┘    └─────────────┘    └─────────────┘    └─────────────┘

Detailed Flow:

  1. Standby: Ava continuously listens for wake word (local processing, no internet)
  2. Wake Detection: When wake word detected, plays prompt sound, starts recording
  3. Audio Transmission: Recording sent to Home Assistant via ESPHome protocol
  4. Speech Recognition: Home Assistant's voice assistant performs speech-to-text
  5. Intent Processing: Home Assistant understands intent and executes action
  6. Speech Synthesis: Home Assistant generates voice response
  7. Playback: Ava receives and plays voice response

Wake Words

Ava supports two independent wake words. You can use one or both simultaneously.

Wake Word Engines

Ava has two on-device wake word engines. Both run entirely locally — no audio leaves the device for wake detection.

microWakeWord (default)

Dimension Value
Architecture TFLite binary classification
Model size 50-80KB, uint8 quantized
Inference 10ms frame, stride 3
Frontend microfeatures
Decision 5-frame sliding window mean > threshold
Output Scalar 0-1 probability
Wake-word swap Full retraining required
CPU / memory Minimal
False wake defense Threshold only (zero-sum trade-off)
Interpretability None
Best for Low-end Android 5+ persistent background

microWakeWord is the default engine. It uses TensorFlow Lite with tiny quantized models. Each wake word is a separate .tflite model file paired with a .json config. The detector runs a sliding window average over the last 5 frames and triggers when the average probability exceeds the cutoff.

Built-in micro models (9):

Model ID Wake Word Author
hey_jarvis Hey Jarvis Kevin Ahrendt
alexa Alexa Kevin Ahrendt
hey_home_assistant Hey Home Assistant Michael Hansen
hey_mycroft Hey Mycroft Kevin Ahrendt
hey_luna Hey Luna adamlonsdale
hey_peppa_pig Hey Peppa Pig Michael Hansen
okay_computer Okay Computer Michael Hansen
okay_nabu OK Nabu Kevin Ahrendt
choo_choo_homie Choo Choo Homie Michael Hansen

Stop word: stop (Stop) by Kevin Ahrendt.

vsWakeWord

Dimension Value
Architecture ONNX CTC phoneme decoding + edit distance
Model size ~500KB
Inference 80ms cycle, 1300ms window, 128×40 feature map
Frontend Log-Mel + adaptive noise floor + spectral VAD
Decision Voice gate → CTC confidence → edit distance ≤1 → 2-hit confirm → 2s cooldown
Output Phoneme sequence with traceability
Wake-word swap Manifest JSON hot-swap
CPU / memory Significantly higher
False wake defense Multi-layer independent gates
Interpretability Phoneme-level debugging
Best for Noise-sensitive, explainability-required deployments

vsWakeWord uses ONNX Runtime with CTC (Connectionist Temporal Classification) decoding. Instead of a binary "is this the wake word?" classifier, it decodes the audio into a phoneme sequence and matches against target phonemes with edit distance ≤1. This makes it more robust to noise and accents, at the cost of higher CPU usage.

Built-in vs models (3):

Model ID Wake Word Type
hey_jarvis Hey Jarvis Wake word
ok_nabu OK Nabu Wake word
ok_stop Ok Stop Stop classifier

vsWakeWord manifest structure:

Each vs model is a pair: id.json (manifest) + id.ort (ONNX model). The manifest defines:

{
  "name": "hey_jarvis",
  "format": "vs-wake-word-ctc-v1",
  "recommended_threshold": 0.61,
  "input": { "shape": [1, 128, 40], "feature": "log_mel" },
  "output": { "shape": [1, 49, 52], "meaning": "frame_level_phoneme_log_probabilities" },
  "feature_config": {
    "sample_rate": 16000, "window_ms": 1300, "frame_ms": 25, "hop_ms": 10,
    "n_fft": 512, "n_mels": 40, "f_min": 80.0, "f_max": 7600.0
  },
  "ctc": {
    "vocab_size": 52, "blank_id": 1, "max_edit_distance": 1,
    "wake_word_targets": [[27, 9, 15, 2, 24, 44, 5, 3, 36, 41, 15, 38]],
    "wake_word_target_phonemes": [["h", "e", "ɪ", " ", "d", "ʒ", "ɑ", "ː", "ɹ", "v", "ɪ", "s"]]
  },
  "runtime": {
    "required_hits": 2, "hit_mode": "consecutive",
    "cooldown_ms": 2000, "high_confidence_bypass": 6.8
  },
  "stop_classifier": false
}

Key fields:

  • wake_word_targets — phoneme ID sequences to match (multiple pronunciations allowed)
  • wake_word_target_phonemes — human-readable phoneme sequences for debugging
  • max_edit_distance — how many phoneme substitutions/insertions/deletions are tolerated (≤1)
  • runtime.required_hits — how many consecutive matches needed to trigger (2 = double confirm)
  • runtime.high_confidence_bypass — skip the hit counter if confidence is very high
  • stop_classifier — true for stop-word models (different gating logic)

Engine Comparison

Dimension microWakeWord vsWakeWord
Architecture TFLite binary classification ONNX CTC phoneme decoding + edit distance
Model size 50-80KB, uint8 quantized ~500KB
Inference 10ms frame, stride 3 80ms cycle, 1300ms window, 128×40 feature map
Frontend microfeatures Adaptive noise floor + spectral VAD
Decision 5-frame sliding window mean Voice gate → CTC confidence → edit distance ≤1 → 2-hit confirm → 2s cooldown
Output Scalar 0-1 probability Phoneme sequence with traceability
Wake-word swap Full retraining required Manifest JSON hot-swap
CPU / memory Minimal Significantly higher
False wake defense Threshold only (zero-sum trade-off) Multi-layer independent gates
Interpretability None Phoneme-level debugging
Best for Low-end Android 5+ persistent background Noise-sensitive, explainability-required deployments

Engine Switching

Each engine stores wake words independently (microWakeWords / vsWakeWords). Switching engines auto-restores the last selection — no more lost models or silent failures.

Cross-engine ID mapping: micro's okay_nabu auto-maps to VS's ok_nabu. HA-configured wake words also resolve correctly across engines.

Service auto-restarts on engine switch, keeping the detector in sync with settings.

Differences from Original brownard/Ava — Wake Word Engine Only

Ava Pro is based on brownard/Ava. This section covers only the wake word engine differences.

Dimension brownard/Ava (original) Ava Pro (knoop7/Ava)
Engine count 1 (microWakeWord) 2 (microWakeWord + vsWakeWord)
microWakeWord engine TFLite binary classification, sliding window threshold Same engine, same 9 built-in models
vsWakeWord engine Not available ONNX CTC phoneme decoding + edit distance + multi-layer false wake gates
Built-in models 9 micro (.tflite) 9 micro (.tflite) + 3 vs (.ort)
Model format (micro) .tflite + .json .tflite + .json (identical, V2/V3 compatible)
Model format (vs) N/A .ort (ONNX) + .json manifest with CTC phoneme targets
Custom model loading DocumentTreeWakeWordProvider (SAF folder picker in Settings) APK assets injection (MT Manager / APK Editor) — see below
False wake defense Threshold only (single layer) Threshold (micro) or voice gate + CTC confidence + edit distance + 2-hit confirm + cooldown (vs)
Inference runtime TensorFlow Lite TensorFlow Lite + ONNX Runtime (reduced build, CPU EP only)
CPU / memory footprint Minimal Minimal (micro) or significantly higher (vs)
Why Ava Pro runs faster than the original

The microWakeWord engine itself is identical between both apps — same models, same inference path. The speed difference comes from what happens around the engine, not inside it.

1. Stop word detection is no longer always-on.

The original runs the stop-word model continuously alongside the wake-word model — two models inference every audio chunk, 24/7, even when nothing is happening. Ava Pro (since 0.5.2) only activates the stop-word model when it is actually useful: when a timer alarm is ringing, or when a voice session is in progress (Listening / Processing / Responding). During idle standby, the stop model is completely skipped. This cuts continuous CPU load in half during the 99% of the time the device is just waiting.

2. vsWakeWord skips inference during silence.

When using the vsWakeWord engine, Ava Pro runs a lightweight voice activity gate before feeding audio to the ONNX model. The gate analyzes audio energy and spectral characteristics in real time. If the input is silence or background noise (no human speech present), the entire ONNX inference is skipped — no model loading, no tensor computation, no phoneme decoding. The ONNX model only wakes up when the gate detects voice-like audio. On a quiet device sitting in a hallway, this means the ONNX engine effectively sleeps most of the time, while still catching the wake word the moment someone speaks.

3. Buffered audio replay on gate open.

When the voice gate transitions from closed to open (someone starts speaking), Ava Pro replays the last few seconds of buffered audio through the ONNX model in a single batch. This means the beginning of the wake word — which happened while the gate was still deciding — is not lost. The user experiences instant wake detection without waiting for the model to "warm up" from silence.

4. Incremental feature extraction.

vsWakeWord extracts log-Mel features from audio. Instead of recomputing the full feature window every chunk, Ava Pro shifts the existing feature buffer and only computes the new frames. On a 1300ms window with 80ms chunks, this means computing ~6 new frames instead of ~128 every cycle — a 20x reduction in FFT work.

5. Adaptive gain normalization.

vsWakeWord applies a smooth adaptive gain to normalize voice volume before inference. This is not about speed directly, but it means the model sees consistently-leveled audio regardless of distance or microphone sensitivity. Consistent input means the CTC confidence scores are more stable, which means the 2-hit confirmation gate reaches its threshold in fewer attempts — faster triggering with fewer false rejects.

6. ONNX Runtime is a stripped build.

The ONNX Runtime shipped with Ava Pro is a custom reduced build — only the CPU execution provider, no GPU/NNAPI delegates, no training APIs. This makes model loading and session creation faster, and the native library is smaller to load into memory. The tradeoff is no hardware acceleration, but for wake-word-sized models the CPU path is already fast enough and avoids the latency and compatibility issues of GPU/NNAPI on fragmented Android devices.

Net effect: On a typical device in idle standby, Ava Pro's wake-word CPU usage is lower than the original because stop-word inference is skipped. When someone speaks, vsWakeWord's voice gate + buffered replay + incremental features make detection feel instant despite the heavier model. The microWakeWord engine path matches the original's speed; the vsWakeWord path trades higher peak CPU for smarter gating and faster perceived response.

Why Ava Pro doesn't use SAF folder picker for custom wake words

The original brownard/Ava ships a DocumentTreeWakeWordProvider that uses Android's Storage Access Framework (SAF). In Settings, you pick a folder, drop .tflite + .json files there, and the app loads them at runtime.

Ava Pro does not include this provider. The reasons are architectural:

  1. Dual engine factory. Ava Pro's WakeWordDetectorFactory dispatches to either microWakeWord or vsWakeWord. The original SAF provider only knows microWakeWord's .tflite format. vsWakeWord uses .ort (ONNX) + manifest JSON with CTC phoneme targets — a completely different model format. A single SAF folder cannot serve both engines without complex format detection and validation logic.

  2. ONNX Runtime safety. TFLite can load arbitrary .tflite files from user storage safely. ONNX Runtime is more sensitive to model format mismatches — an invalid .ort file could crash the native inference session or cause OOM on low-end devices. The asset-bundled approach guarantees models are validated at build time.

  3. Build-time validation. By bundling models in assets/, the build system catches format errors, missing manifest fields, and phoneme inventory mismatches before the app ships. SAF-loaded models have no such guarantee — a malformed JSON could silently disable wake detection.

  4. Scope. Ava Pro added vsWakeWord, dual wake words, voiceprint, visual feedback, and stop word optimization. Re-implementing SAF support for both engine formats safely was deprioritized in favor of these features.

How to add custom wake words to Ava Pro (APK injection, no source build required)

Ava Pro loads microWakeWord models from assets/wakeWords/ and vsWakeWord models from assets/vswakeword/ inside the APK. You can inject custom models without building from source — just edit the APK directly.

Tools needed:

  • Android file manager with APK editing: MT Manager (Chinese), APK Editor Pro, or Nexus APK Editor
  • Or on desktop: apktool + zipalign + apksigner

Steps (MT Manager method):

  1. Download the Ava Pro lite APK from GitHub releases
  2. Open MT Manager, long-press the APK, select "View" (or "Extract")
  3. Navigate to assets/wakeWords/ (for microWakeWord) or assets/vswakeword/ (for vsWakeWord)
  4. Copy your custom model files into the directory:
    • microWakeWord: my_word.tflite + my_word.json
    • vsWakeWord: my_word.ort + my_word.json
  5. Save and repack the APK (MT Manager handles re-signing automatically)
  6. Uninstall the old Ava, install the modified APK
  7. Open Settings -> Voice Config -> Wake Word, your custom model appears in the list

Steps (desktop apktool method):

# Decompile
apktool d Ava-0.5.4.apk -o ava_decoded

# Add your model
cp my_word.tflite my_word.json ava_decoded/assets/wakeWords/

# For vsWakeWord
cp my_word.ort my_word.json ava_decoded/assets/vswakeword/

# Repack
apktool b ava_decoded -o Ava-custom.apk

# Sign
zipalign -v 4 Ava-custom.apk Ava-custom-aligned.apk
apksigner sign --ks ava-key.jks --ks-pass pass:1234 --out Ava-custom-signed.apk Ava-custom-aligned.apk

JSON format requirements:

For microWakeWord (same as V3 community format):

{
  "type": "micro",
  "wake_word": "My Custom Word",
  "author": "Your Name",
  "model": "my_word.tflite",
  "trained_languages": ["en"],
  "version": 2,
  "micro": {
    "probability_cutoff": 0.6,
    "sliding_window_size": 5,
    "feature_step_size": 10,
    "tensor_arena_size": 30000,
    "minimum_esphome_version": "2024.7.0"
  }
}

For vsWakeWord (see manifest structure above for full template).

Note: This is an advanced operation intended for power users. Wake word customization is inherently an advanced topic — the models require training or downloading from trusted sources. The APK injection method works because Ava Pro reads models from the APK's assets/ directory at runtime, same as the built-in models. No source code compilation is needed.

Compatibility: Both apps share the same microWakeWord model format (V2/V3 .tflite + .json). Models from TaterTotterson/microWakeWords V3 directory work directly. Just drop the .tflite + .json pair into assets/wakeWords/.

Custom Wake Words

microWakeWord: Download Pretrained or Train Your Own

Ava uses microWakeWords V3 format. The JSON config is identical to the V3 models from the community.

Option 1: Download Pretrained Models (Easiest)

The microWakeWords repo by TaterTotterson maintains a large library of pretrained V3 models. The format is directly compatible with Ava.

  1. Browse the microWakeWordsV3 directory
  2. Find a wake word you like (e.g., aleesa, angel, annika, arale, artamis, etc.)
  3. Download both files: name.json and name.tflite
  4. Place them in Ava's app/src/main/assets/wakeWords/ directory
  5. Rebuild the APK

The V3 JSON format is identical to Ava's built-in models:

{
  "type": "micro",
  "wake_word": "ah_lehks_sah",
  "author": "Tater Totterson",
  "website": "https://github.com/TaterTotterson/microWakeWord-Trainer-AppleSilicon",
  "model": "ah_lehks_sah.tflite",
  "trained_languages": ["en"],
  "version": 2,
  "micro": {
    "probability_cutoff": 0.1,
    "sliding_window_size": 3,
    "feature_step_size": 10,
    "tensor_arena_size": 30000,
    "minimum_esphome_version": "2024.7.0"
  }
}
Option 2: Train Your Own with macOS Trainer (Apple Silicon)

If you have a Mac with Apple Silicon (M1/M2/M3/M4), you can train a custom wake word with a local web UI:

  1. Install the trainer:

    • Download the signed macOS app from WakeWord Trainer releases
    • Or clone and run from source: 
      git clone https://github.com/TaterTotterson/microWakeWord-Trainer-AppleSilicon.git
cd microWakeWord-Trainer-AppleSilicon
./run.sh
    • Open http://127.0.0.1:8789 in your browser

  2. Train the wake word:

    • Enter your wake phrase in the Trainer tab
    • Choose language (en, or other Piper-supported languages)
    • Optionally test pronunciation with Test TTS
    • Click Start training
    • The trainer uses Piper TTS to generate samples automatically
    • Personal samples are optional but improve accuracy

  3. Optionally capture real samples from devices:

    • Flash a device with Tater firmware (from the Firmware tab)
    • Enable Capture Wake Audio on the device
    • Set Trainer App URL to http://<trainer-ip>:8789
    • Review captured clips in the Captured Audio tab
    • Mark good clips as "This is good", bad ones as "False wake"

  4. Get the output files:

    • Successful training produces:
      • trained_wake_words/<wake_word>.tflite
      • trained_wake_words/<wake_word>.json

  5. Install into Ava:

    • Copy both files to Ava's app/src/main/assets/wakeWords/ directory
    • Rebuild the APK
Option 3: Train with Direct Script
git clone https://github.com/TaterTotterson/microWakeWord-Trainer-AppleSilicon.git
cd microWakeWord-Trainer-AppleSilicon
./train_microwakeword_macos.sh "hey_my_custom_word"

If personal_samples/*.wav or negative_samples/*.wav exist in the folder, they are included automatically.

JSON Config Parameters
Parameter Description Typical Value
probability_cutoff Detection threshold (lower = easier to trigger, more false positives) 0.1 - 0.97
sliding_window_size Frames to average before triggering 3 - 9
feature_step_size Feature extraction step in ms 10
tensor_arena_size TFLite arena size in bytes (must match model) 21000 - 30000
minimum_esphome_version Minimum ESPHome version 2024.7.0

Note: microWakeWord requires a trained model for each wake word. There is no hot-swap — you need a .tflite file. But the V3 community library has 100+ pretrained models you can download directly.

vsWakeWord: Manifest Hot-Swap

vsWakeWord supports manifest JSON hot-swap. To create a custom wake word:

  1. Train a CTC model using the vsWakeWord training pipeline (PyTorch → ONNX export)
  2. Create a manifest JSON with your phoneme targets: 
    {
  "name": "my_custom_word",
  "format": "vs-wake-word-ctc-v1",
  "recommended_threshold": 0.6,
  "input": { "name": "input", "shape": [1, 128, 40], "dtype": "float32", "feature": "log_mel" },
  "output": { "name": "log_probs", "shape": [1, 49, 52], "dtype": "float32" },
  "feature_config": {
    "sample_rate": 16000, "window_ms": 1300, "frame_ms": 25, "hop_ms": 10,
    "n_fft": 512, "n_mels": 40, "f_min": 80.0, "f_max": 7600.0, "log_floor": 1e-06
  },
  "ctc": {
    "vocab_size": 52, "blank_id": 1, "pad_id": 0, "word_sep_id": 2,
    "wake_word_targets": [[your_phoneme_ids]],
    "wake_word_target_phonemes": [["your", "phonemes"]],
    "max_edit_distance": 1
  },
  "runtime": {
    "required_hits": 2, "hit_mode": "consecutive",
    "cooldown_ms": 2000, "high_confidence_bypass": 6.8
  },
  "stop_classifier": false
}
  3. Place both files (id.json + id.ort) in assets/vswakeword/ directory
  4. Rebuild the Ava APK

The 52-phoneme inventory uses IPA-style symbols. See an existing manifest (e.g., hey_jarvis.json) for the full inventory list.

Advantage: vsWakeWord's manifest-based approach lets you swap wake word targets without retraining the base model in some cases — just update the wake_word_targets phoneme IDs. However, for best accuracy, a model trained on your specific wake word is recommended.

How to Change Wake Word

  1. Open Ava app
  2. Go to Settings -> Voice Config
  3. Find Wake Word Engine and choose microWakeWord or vsWakeWord
  4. Find Wake Word 1 option
  5. Select your preferred wake word from the list
  6. Optionally configure Wake Word 2 for dual wake word mode
  7. New wake word takes effect after service restart (auto)

Wake Word Sensitivity

Adjust sensitivity to control how easily the wake word triggers:

  • Higher sensitivity = easier to trigger, but more false positives
  • Lower sensitivity = fewer false positives, but may miss quiet speech

Wake Sound

Each wake word can have its own wake sound:

  • Wake Word 1 Sound: Played when Wake Word 1 is detected
  • Wake Word 2 Sound: Played when Wake Word 2 is detected
  • Default Sound: Used if no custom sound is set
  • None: Silent recording start

Wake Visual Feedback

Ava provides clear visual feedback during wake and conversation:

Wake Instant:

  • Colorful ripple expanding from screen center
  • Android 13+: RuntimeShader with distorted halo + star particles
  • Android 7: Soft circular diffusion
  • Compatibility paths for other versions

Conversation (when Floating Subtitle is disabled):

  • Full-screen edge glow that changes with state:
    • Listening: Edge light breathes with microphone volume
    • Processing: Slow breathing animation
    • Speaking: Pulsates with TTS energy

Dual Wake-Word Color Coding:

  • Wake Word 1 = green (default)
  • Wake Word 2 = blue (default)
  • Ripple and edge light match the triggered wake word
  • Custom colors available in Settings → Extensions → Interface → Voice feedback colors
  • 7 rainbow presets (red through purple) also available

Technical Notes:

  • Edge glow uses pre-rendered Gaussian blur bitmaps for performance
  • Ripple animation driven by system uptime (prevents Kiosk devices with "animation duration = 0" from killing the effect)
  • Android 7.0/7.1 optimized to clean circular diffusion without Shader dependency

Voiceprint Recognition

Ava Pro includes an experimental on-device smart voiceprint recognition feature. Once enabled, Ava learns frequent wakeword callers from short local samples. On subsequent wakeword triggers, it identifies who likely woke the device and publishes the result to a Home Assistant "Voiceprint Status" sensor.

This is not a heavyweight cloud voiceprint model. It is a lightweight local matcher purpose-built for Ava's fixed-wakeword scenario. It uses short local samples, audio feature extraction, quality gating, and conservative learning — no large model packages required.

How It Works

  • After you wake the device a few times, it learns who usually speaks
  • All audio stays local, nothing is uploaded to cloud
  • Two user slots available, each with a custom display name

Setup

  1. Go to Settings -> Voice Config
  2. Turn on Smart voiceprint recognition
  3. Set User 1 name and User 2 name
  4. Wake the device a few times for each user
  5. The voiceprint status sensor appears in Home Assistant

Home Assistant Entity

sensor.your_device_name_voice_print_status

Audio Configuration

System Recording Mode

Choose how the microphone captures audio:

Mode Description
Auto select Lets the device choose. Best default
Speech boost Tries to make speech clearer. May change tone
Normal pickup Plain microphone input. Safest choice
Call boost More like call audio. May over-process
Unprocessed Closest to raw input. Good for testing

Audio Processing

Option Description
Noise suppression Use device noise suppressor when available
Automatic gain control Use device AGC when available
Hardware echo cancellation Enable hardware echo cancellation when supported
Software echo cancellation Cancel Ava's own playback from microphone so wake words work during music/TTS playback

Software AEC is recommended when using music playback or TTS. It replaces the device echo canceler while enabled.

Software Mic Gain

Boost PCM audio before wake word detection and streaming. Range: 0-24 dB.

Audio Config Presets

Preset Description
Standard 16kHz Mono Default for voice recognition
Broadcast 48kHz Mono Higher quality capture
Stereo Input 48kHz Stereo recording
Low Latency 16kHz Reduced latency
Voice Call Optimized For voice call use case
Unprocessed Raw 48kHz Raw capture, no processing
Compact 8kHz Lowest bandwidth
CD Quality 44.1kHz High quality
Studio 96kHz Very high quality
Ultra HD 192kHz Maximum quality

Voice Channel

The Voice Channel switch disables all voice input, wake word, and voice assistant threads. Service restart required to apply.

Stop Words

Stop words interrupt the current conversation or stop Ava's response (e.g., timer alarm).

Stop Word Engine Model ID Description
Stop microWakeWord stop Default stop word for micro engine
Ok Stop vsWakeWord ok_stop Stop word for vs engine (stop_classifier: true)

Stop-word detection only runs when actually needed (since 0.5.2):

  1. A timer alarm is actively ringing
  2. A voice session is in progress (Listening, Processing, Responding)

During idle standby with no alarm, the stop model is skipped — cutting CPU load and heat.

Continuous Conversation

Continuous conversation lets you issue multiple commands without saying the wake word each time.

How It Works

  1. Say wake word + first command
  2. After Ava responds, automatically enters listening mode
  3. Say next command directly (no wake word needed)
  4. After 10 seconds of silence, exits continuous conversation mode

End Conditions

Mode Description
Exit Keyword Stop End when the assistant says goodbye or farewell phrases (default)
Question Mark Continue Keep listening only if the reply ends with "?"; stop after other replies

How to Enable

  1. Go to Settings -> Interaction
  2. Turn on Continuous Conversation switch
  3. Choose end condition mode

Conversation Subtitles

Conversation subtitles display what you said and Ava's response on screen as a floating window.

How to Enable

  1. Go to Settings -> Interaction
  2. Turn on Floating Subtitle switch

Mute Mode

Mute mode turns off the microphone, Ava won't respond to any wake words.

How to Enable

Method 1: In Settings

  1. Go to Settings -> Voice Config
  2. Turn on Mute switch

Method 2: Home Assistant Control

service: switch.turn_on
target:
  entity_id: switch.your_device_name_mute

Settings Summary

Setting Location Description Default
Device Name Voice Config Name shown in HA device_model_voice_assistant
Port Voice Config ESPHome communication port 6053
Wake Word Engine Voice Config microWakeWord or vsWakeWord microWakeWord
Wake Word 1 Voice Config Primary wake word Hey Jarvis
Wake Word 2 Voice Config Secondary wake word None
Sensitivity Voice Config Wake word sensitivity -
Wake Sound Voice Config Prompt sound when recording starts Optional
Voice Channel Voice Config Master voice input switch On
System Recording Mode Voice Config Audio capture mode Auto select
Noise suppression Voice Config Device noise suppressor -
Automatic gain control Voice Config Device AGC -
Hardware echo cancellation Voice Config Hardware AEC -
Software echo cancellation Voice Config Software AEC for playback -
Voiceprint recognition Voice Config Speaker recognition Off
Software Mic Gain Voice Config PCM boost 0-24 dB 0
Audio Config Voice Config Capture preset Standard 16kHz Mono
Mute Voice Config Turn off microphone Off
Continuous Conversation Interaction Issue commands continuously Off
Floating Subtitle Interaction Display conversation text Off

Home Assistant Services

Manually Trigger Wake

service: esphome.your_device_name_trigger_wake
data: {}

Control Mute

# Enable mute
service: switch.turn_on
target:
  entity_id: switch.your_device_name_mute

# Disable mute
service: switch.turn_off
target:
  entity_id: switch.your_device_name_mute

Set Volume

service: media_player.volume_set
target:
  entity_id: media_player.your_device_name
data:
  volume_level: 0.8  # 0.0 - 1.0

Send Voice Command (Text)

service: text.set_value
target:
  entity_id: text.your_device_name_voice_command
data:
  value: "Turn on the living room light"

FAQ

Ava can't hear me?

  1. Check if microphone permission is granted
  2. Check if mute mode is enabled
  3. Check if Voice Channel is enabled
  4. Make sure device volume isn't muted
  5. Try speaking closer to device
  6. Try a different System Recording Mode

Wake word recognition inaccurate?

  1. Try a different wake word
  2. Adjust wake word sensitivity
  3. Make sure environment isn't too noisy
  4. Enable noise suppression
  5. Speak at moderate speed, pronounce clearly
  6. Try vsWakeWord engine — its CTC phoneme matching is more robust to noise and accents

Wake word doesn't work during music playback?

  1. Enable Software echo cancellation in Settings -> Voice Config
  2. This cancels Ava's own playback from the microphone input

Speech recognition results wrong?

This is usually a Home Assistant issue:

  1. Check HA voice assistant configuration
  2. Make sure Whisper etc. components are working
  3. Check network latency
  4. Try speaking more clearly

Back to Home

Add a custom footer

Core

Optional Modules

Advanced

Clone this wiki locally