Bump ESPHome pin to 2026.5.1

Added additional images and a link to the README for better presentation.

.github/workflows/docker-publish.yml

@@ -0,0 +1,48 @@
+name: Publish Docker Image
+
+on:
+  push:
+    branches:
+      - main
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  packages: write
+
+concurrency:
+  group: docker-publish-${{ github.ref }}
+  cancel-in-progress: true
+
+env:
+  REGISTRY: ghcr.io
+  IMAGE_NAME: tatertotterson/microwakeword
+
+jobs:
+  docker:
+    name: Docker image
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to GHCR
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Build and push image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: dockerfile
+          platforms: linux/amd64
+          push: true
+          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest
+          cache-from: type=gha,scope=mww-trainer-nvidia-docker
+          cache-to: type=gha,mode=max,scope=mww-trainer-nvidia-docker

.gitignore

@@ -1,3 +1,4 @@
 personal_samples/*
 data/
+trim_history/
 .DS_Store

README.md

@@ -1,16 +1,15 @@
 <div align="center">
-  <h1>microWakeWord NVIDIA Docker Trainer UI</h1>
+  <a href="https://taterassistant.com">
-  <img width="800" alt="Screenshot 2026-04-14 at 11 02 06 PM" src="https://github.com/user-attachments/assets/694f4cb7-e4d8-4e2b-80ec-b40fb41cbfff" />
+    <img src="images/tater-repo-logo.png" alt="microWakeWord Trainer" width="460"/>
+  </a>
 </div>
+<h3 align="center">
+  <a href="https://taterassistant.com">taterassistant.com</a>
+</h3>
 
-Train custom microWakeWord models in Docker with:
+Train custom microWakeWord models in Docker with NVIDIA/CUDA acceleration, generated Piper samples, device-captured samples, reviewed false-wake negatives, live training logs, and ESPHome firmware flashing.
 
-- uploaded personal voice samples
+Real samples come from device-captured wake audio, close misses, or manual uploads. Every saved sample is normalized to `16 kHz / mono / 16-bit PCM WAV` before training.
-- automatically generated Piper TTS samples
-- a browser-based trainer UI
-- live training logs in a popup console
-
-This project no longer records audio in the browser. The UI is now upload-first: users add their own audio files, the app validates or converts them, and training runs from the same page.
 
 ---
 
@@ -27,41 +26,92 @@ docker pull ghcr.io/tatertotterson/microwakeword:latest
 ```bash
 docker run -d \
   --gpus all \
-  -p 8888:8888 \
+  --network host \
+  -e REC_PORT=8789 \
   -v $(pwd):/data \
   ghcr.io/tatertotterson/microwakeword:latest
 ```
 
-What these flags do:
+The flags:
 
-- `--gpus all` enables GPU acceleration
+- `--gpus all` enables GPU acceleration.
-- `-p 8888:8888` exposes the trainer UI
+- `--network host` lets the container receive mDNS/zeroconf traffic for ESPHome auto-detect.
-- `-v $(pwd):/data` persists models, downloaded voices, datasets, and personal samples
+- `-e REC_PORT=8789` sets the trainer web UI and captured-audio port. Change this value if `8789` is already in use.
+- `-v $(pwd):/data` persists models, downloaded voices, datasets, samples, and firmware caches.
 
-Then open:
+Host networking is recommended for the Firmware tab's mDNS device discovery. Manual IP flashing and captured-audio uploads can still work without host networking if the trainer port is reachable, but auto-detect may not see devices from Docker bridge networking.
+
+Open:
 
 ```text
-http://localhost:8888
+http://localhost:8789
 ```
 
+If you change `REC_PORT`, open that port instead and use the same port in the ESPHome `Trainer App URL`.
+
 ---
 
 ## What The UI Does
 
-- Start a wake word session
+- `Trainer` starts a wake-word session, shows positive/negative sample counts, and launches training.
-- Test TTS pronunciation
+- `Captured Audio` reviews clips sent by ESPHome sats, including wake hits, close misses, and false wakes.
-- Upload one or many personal samples
+- `Samples` plays, removes, clears, and manually imports personal or negative samples.
-- Normalize uploads to `16 kHz / mono / 16-bit PCM WAV`
+- `Firmware` builds the latest `microWakeWords` ESPHome YAMLs from GitHub and flashes VoicePE or Satellite1 over OTA.
-- Train with or without personal samples
+- Popup consoles show colorized training and firmware logs while long-running jobs are active.
-- Show a popup console with live progress and logs
-
-Personal samples are optional. If none are uploaded, the trainer can still proceed with TTS-only data after confirmation.
 
 ---
 
-## Personal Samples
+## Captured Audio Workflow
 
-Accepted upload formats include:
+To collect samples from a sat, flash it with the Tater firmware from [TaterTotterson/microWakeWords](https://github.com/TaterTotterson/microWakeWords). The `Firmware` tab can build and flash the VoicePE or Satellite1 YAMLs directly from that repo.
+
+After flashing, the device exposes ESPHome entities for capture setup:
+
+- `Capture Wake Audio` toggles upload of wake-word triggers.
+- `Capture Close Misses` toggles upload of near misses.
+- `Trainer App URL` sets the trainer address, for example `http://<trainer-ip>:8789`.
+
+ESPHome devices can send raw captured audio to:
+
+```text
+/api/upload_captured_audio_raw
+```
+
+Keep the training app running and reachable at the `Trainer App URL` while capture is enabled. The sats upload clips live; if the app is stopped or the URL is wrong, captured audio will not be saved.
+
+In the `Captured Audio` tab:
+
+- play each clip from the inbox
+- mark good wake-word clips as `This is good`
+- mark bad triggers as `False wake`
+- discard clips that should not be used
+
+Approved clips move into:
+
+```text
+/data/personal_samples/
+```
+
+False wakes move into:
+
+```text
+/data/negative_samples/
+```
+
+Captured audio is boosted for easier playback in the UI, then kept in the correct training format.
+
+---
+
+## Samples
+
+The `Samples` tab is the sample library.
+
+- `Personal` samples are positive examples of the wake word.
+- `Negative` samples are reviewed false wakes or hard negatives.
+- Both can be played back and removed one at a time.
+- Manual upload is available here as an optional seed path.
+
+Accepted manual upload formats include:
 
 - WAV
 - MP3
@@ -72,97 +122,120 @@ Accepted upload formats include:
 - OPUS
 - WEBM
 
-The backend validates or converts uploads with `ffmpeg` and stores the normalized files in:
+Uploads are validated or converted with `ffmpeg` into:
 
 ```text
-/data/personal_samples/
+16 kHz / mono / 16-bit PCM WAV
 ```
 
-Notes:
+Starting a new session does not clear samples. Use the clear buttons in `Samples` if you want to remove saved personal or negative clips.
 
-- starting a new session does not clear personal samples
+---
-- use the `Clear personal samples` button if you want to wipe them
+
-- any uploaded personal samples are automatically included in training
+## Training Flow
+
+1. Enter the wake phrase in `Trainer`.
+2. Choose the language.
+3. Optionally test pronunciation with `Test TTS`.
+4. Review the positive and negative sample counts.
+5. Click `Start training`.
+6. Watch the popup training console.
+
+Personal samples are optional. Training can run with zero personal samples after confirmation, using generated TTS samples and the stock negative datasets.
+
+Reviewed negative samples are converted into `/data/work/reviewed_negative_features/` and inserted into the training YAML as a hard-negative feature set when present.
 
 ---
 
 ## Language Support
 
-The language selector is dynamic.
+The language picker is dynamic.
 
-- `en` is always available
+- `en` is always available.
-- non-English languages are populated from Piper voice metadata
+- English keeps the existing dedicated generator model path.
-- when you train with a non-English language, the backend downloads all Piper ONNX voices for that selected language only
+- Non-English languages are discovered from the Piper voices catalog and any local Piper voice metadata.
-- it does not pre-download every language
+- When a non-English language is selected, the trainer downloads all voices for that selected language only.
-- already-downloaded voices are reused on later runs
+- Already-downloaded voices are reused.
+- It does not download every language up front.
 
-English stays on its existing dedicated generator model path. Non-English languages use the selected language's ONNX Piper voices.
+If the upstream Piper catalog is unavailable, already-installed local voices are used when available.
-
-If the Piper catalog is unavailable, already-installed local voices can still be used.
 
 ---
 
-## Training Behavior
+## Dataset Behavior
 
-1. Enter the wake word
+The first training run downloads and prepares missing training assets into `/data`, including:
-2. Optionally test pronunciation
-3. Optionally upload personal samples
-4. Click `Start training`
-5. Watch the popup console for:
-   - selected-language voice downloads when needed
-   - sample generation progress
-   - dataset setup
-   - training progress and completion
-
-The `Open console` button lets you reopen the log window after closing it.
-
----
-
-## First Run Notes
-
-The first real training run may download large training assets into `/data`, such as:
 
 - Piper voices for the selected language
-- training datasets and background data
+- negative datasets and background data
-- Python training environment dependencies
+- the Python training environment
+- generated samples and augmented feature caches
 
-These are reused later unless you delete `/data`.
+After those assets are prepared, later runs reuse the local copies unless the mounted `/data` contents are deleted.
+
+---
+
+## Firmware Flashing
+
+The `Firmware` tab builds and flashes Tater firmware for supported ESPHome sats.
+
+- Downloads the latest firmware YAML templates from `TaterTotterson/microWakeWords` on GitHub.
+- Lets you choose `VoicePE` or `Satellite1`.
+- Auto-detects ESPHome devices with mDNS when the container is running with host networking.
+- Allows manual IP or hostname entry if discovery does not find the device.
+- Saves firmware form values so you do not re-enter sounds and URLs every run.
+- Lists locally trained wake words from `/data/trained_wake_words/` for easy model selection.
+- Builds with ESPHome and flashes OTA.
+- Streams ESPHome output in a colorized firmware console.
+
+Firmware YAMLs are intentionally pulled from GitHub each time. There is no local fallback path in the trainer UI.
 
 ---
 
 ## Output Files
 
-Successful runs produce:
+Successful runs produce timestamped training output folders such as:
 
 ```text
-/data/output/<wake_word>.tflite
+/data/output/<timestamp>-<wake_word>-<samples>-<steps>/<wake_word>.tflite
-/data/output/<wake_word>.json
+/data/output/<timestamp>-<wake_word>-<samples>-<steps>/<wake_word>.json
 ```
 
-If those files already exist, the trainer creates timestamped backups before replacing them.
+The trainer also syncs firmware-ready artifacts into:
+
+```text
+/data/trained_wake_words/<wake_word>.tflite
+/data/trained_wake_words/<wake_word>.json
+```
+
+The firmware tab uses `/data/trained_wake_words/` to populate the wake-word dropdown.
 
 ---
 
 ## Resetting Everything
 
-If you want a clean slate, stop the container and remove the contents of your mounted `/data` directory.
+If you want a clean slate, stop the container and remove the contents of the mounted `/data` directory.
 
-That will remove:
+That removes:
 
 - personal samples
+- negative samples
+- captured inbox clips
 - downloaded Piper voices
 - cached datasets
 - training environments
 - trained models
+- firmware build caches
 
 ---
 
-## Notes
+## Important Notes
 
-- browser microphone recording has been removed
+- Personal samples are optional.
-- personal samples are optional
+- Negative samples are optional but useful for reducing false wakes.
-- the server module is now `trainer_server.py`
+- The UI server is `trainer_server.py`.
-- the launcher script is now `run.sh`
+- The launcher is `run.sh`.
+- Firmware capture settings live on the ESPHome device and can be toggled from the device entities after flashing.
 
 ---
 

cli/wake_word_sample_augmenter

@@ -18,6 +18,8 @@ parser.add_argument("--output-dir", type=str, help="Wake word output dir. Defaul
 # Personal inputs/outputs (NEW)
 parser.add_argument("--personal-dir", type=str, help="Personal WAV dir. Default: <data-dir>/personal_samples", required=False)
 parser.add_argument("--personal-output-dir", type=str, help="Personal features output dir. Default: <data-dir>/work/personal_augmented_features", required=False)
+parser.add_argument("--negative-dir", type=str, help="Reviewed negative WAV dir. Default: <data-dir>/negative_samples", required=False)
+parser.add_argument("--negative-output-dir", type=str, help="Reviewed negative features output dir. Default: <data-dir>/work/reviewed_negative_features", required=False)
 
 # Dataset dirs
 parser.add_argument("--mit-rirs-16k-dir", type=str, help="MIT RIR input directory. Default: <data-dir>/training_datasets/mit_rirs_16k", required=False)
@@ -57,6 +59,17 @@ if not args.personal_output_dir:
 else:
     args.personal_output_dir = os.path.realpath(args.personal_output_dir)
 
+# Reviewed negative defaults
+if not args.negative_dir:
+    args.negative_dir = os.path.join(args.data_dir, "negative_samples")
+else:
+    args.negative_dir = os.path.realpath(args.negative_dir)
+
+if not args.negative_output_dir:
+    args.negative_output_dir = os.path.join(work_dir, "reviewed_negative_features")
+else:
+    args.negative_output_dir = os.path.realpath(args.negative_output_dir)
+
 # Dataset defaults
 if not args.mit_rirs_16k_dir:
     args.mit_rirs_16k_dir = os.path.join(args.data_dir, "training_datasets", "mit_rirs_16k")
@@ -205,7 +218,7 @@ def bind_wav_generator(clips_obj: Clips, wav_dir: str):
 
     clips_obj.audio_generator = types.MethodType(audio_generator_from_wavs, clips_obj)
 
-def generate_feature_set(input_wav_dir: str, out_root_dir: str, label: str):
+def generate_feature_set(input_wav_dir: str, out_root_dir: str, label: str, *, remove_silence: bool = True):
     files = glob.glob(os.path.join(input_wav_dir, "*.wav"))
     if not files:
         print(f"ℹ️  No WAVs found for {label} in: {input_wav_dir} (skipping)")
@@ -218,7 +231,7 @@ def generate_feature_set(input_wav_dir: str, out_root_dir: str, label: str):
         input_directory=input_wav_dir,
         file_pattern="*.wav",
         max_clip_duration_s=5,
-        remove_silence=True,
+        remove_silence=remove_silence,
         random_split_seed=10,
         split_count=0.1,
     )
@@ -263,9 +276,12 @@ def generate_feature_set(input_wav_dir: str, out_root_dir: str, label: str):
 # Wake word generated/TTS features (existing behavior)
 generate_feature_set(args.input_dir, args.output_dir, "generated")
 
-# Personal features (NEW)
+# Personal features
 generate_feature_set(args.personal_dir, args.personal_output_dir, "personal")
 
+# Reviewed false-positive / hard-negative features
+generate_feature_set(args.negative_dir, args.negative_output_dir, "reviewed negatives", remove_silence=False)
+
 END_TIME = datetime.now(timezone.utc).replace(microsecond=0)
 et = END_TIME - START_TIME
 print(f"\n{'=' * 80}")

cli/wake_word_sample_trainer

@@ -111,6 +111,16 @@ else
   echo "ℹ️  No personal features found at ${PERSONAL_FEATURES_DIR}/training (continuing without personal weighting)"
 fi
 
+# Reviewed false-positive features are optional hard negatives.
+REVIEWED_NEGATIVE_FEATURES_DIR="${WORK_DIR}/reviewed_negative_features"
+HAS_REVIEWED_NEGATIVE="false"
+if [ -d "${REVIEWED_NEGATIVE_FEATURES_DIR}/training" ] ; then
+  HAS_REVIEWED_NEGATIVE="true"
+  echo "✅ Found reviewed negative features: ${REVIEWED_NEGATIVE_FEATURES_DIR}/training (will weight as hard negatives)"
+else
+  echo "ℹ️  No reviewed negative features found at ${REVIEWED_NEGATIVE_FEATURES_DIR}/training (continuing with stock negatives)"
+fi
+
 cd "${WORK_DIR}"
 
 echo "===== Starting ${TRAINING_STEPS} training steps ====="
@@ -133,6 +143,7 @@ features:
   truth: true
   type: mmap
 __PERSONAL_FEATURE_MARKER__
+__REVIEWED_NEGATIVE_FEATURE_MARKER__
 - features_dir: __NEG_SPEECH__
   penalty_weight: 1.0
   sampling_weight: 12.0
@@ -208,6 +219,22 @@ else
   sed -i -e "/__PERSONAL_FEATURE_MARKER__/d" "${YAML_PATH}"
 fi
 
+# Insert/remove reviewed hard-negative block
+if [ "${HAS_REVIEWED_NEGATIVE}" = "true" ]; then
+  reviewed_negative_block="$(cat <<EOF
+- features_dir: ${REVIEWED_NEGATIVE_FEATURES_DIR}
+  penalty_weight: 1.25
+  sampling_weight: 8.0
+  truncation_strategy: random
+  truth: false
+  type: mmap
+EOF
+)"
+  perl -0777 -i -pe "s#__REVIEWED_NEGATIVE_FEATURE_MARKER__#${reviewed_negative_block}#g" "${YAML_PATH}"
+else
+  sed -i -e "/__REVIEWED_NEGATIVE_FEATURE_MARKER__/d" "${YAML_PATH}"
+fi
+
 echo "   Wrote training_parameters.yaml"
 rm -rf "${WORK_DIR}/trained_models/wakeword"
 

dockerfile

@@ -6,7 +6,7 @@ ENV DEBIAN_FRONTEND=noninteractive
 # System deps
 RUN apt-get update && apt-get install -y --no-install-recommends \
     python3.12 python3.12-venv python3.12-dev python3-pip python-is-python3 \
-    git wget curl unzip ca-certificates nano less \
+    git wget curl unzip patch ninja-build ca-certificates nano less \
  && rm -rf /var/lib/apt/lists/* \
  && mkdir -p /data
 

run.sh

@@ -6,7 +6,7 @@ ROOTDIR="$(dirname "$(realpath "$0")")"
 # Training convention
 DATA_DIR="${DATA_DIR:-/data}"
 HOST="${REC_HOST:-0.0.0.0}"
-PORT="${REC_PORT:-8888}"
+PORT="${REC_PORT:-8789}"
 
 # Keep trainer UI deps separate from the training venv
 VENV_DIR="${DATA_DIR}/.recorder-venv"
@@ -17,6 +17,7 @@ PIN_FILE="${VENV_DIR}/.pinned_installed"
 FASTAPI_VERSION="${REC_FASTAPI_VERSION:-0.115.6}"
 UVICORN_VERSION="${REC_UVICORN_VERSION:-0.30.6}"
 PY_MULTIPART_VERSION="${REC_PY_MULTIPART_VERSION:-0.0.9}"
+ESPHOME_VERSION="${REC_ESPHOME_VERSION:-2026.5.1}"
 
 echo "microWakeWord Trainer UI (Docker)"
 echo "-> ROOTDIR:  ${ROOTDIR}"
@@ -25,6 +26,16 @@ echo "-> URL:      http://localhost:${PORT}/"
 
 mkdir -p "${DATA_DIR}"
 
+install_ui_deps() {
+  ${PIP} install \
+    "fastapi==${FASTAPI_VERSION}" \
+    "uvicorn[standard]==${UVICORN_VERSION}" \
+    "python-multipart==${PY_MULTIPART_VERSION}" \
+    "esphome==${ESPHOME_VERSION}" \
+    "silero-vad>=5.0.0" \
+    "numpy>=1.24.0"
+}
+
 # -----------------------------
 # Trainer UI venv (separate)
 # -----------------------------
@@ -39,21 +50,63 @@ source "${VENV_DIR}/bin/activate"
 if [[ ! -f "${PIN_FILE}" ]]; then
   echo "Installing pinned trainer UI deps"
   ${PIP} install -U pip setuptools wheel
-  ${PIP} install \
+  install_ui_deps
-    "fastapi==${FASTAPI_VERSION}" \
-    "uvicorn[standard]==${UVICORN_VERSION}" \
-    "python-multipart==${PY_MULTIPART_VERSION}"
   touch "${PIN_FILE}"
 else
   echo "Reusing existing trainer UI venv (no upgrades)"
-fi
+  if ! "${PY}" - "${FASTAPI_VERSION}" "${UVICORN_VERSION}" "${PY_MULTIPART_VERSION}" "${ESPHOME_VERSION}" <<'PY' >/dev/null 2>&1
+import importlib.metadata as md
+import sys
 
+fastapi_version, uvicorn_version, multipart_version, esphome_version = sys.argv[1:5]
+
+def version_tuple(value):
+    parts = []
+    for token in str(value).replace("-", ".").split("."):
+        if token.isdigit():
+            parts.append(int(token))
+        else:
+            digits = "".join(ch for ch in token if ch.isdigit())
+            if digits:
+                parts.append(int(digits))
+            break
+    return tuple(parts)
+
+exact = {
+    "fastapi": fastapi_version,
+    "uvicorn": uvicorn_version,
+    "python-multipart": multipart_version,
+    "esphome": esphome_version,
+}
+minimum = {
+    "silero-vad": "5.0.0",
+    "numpy": "1.24.0",
+}
+present = ("torch", "zeroconf")
+
+for package, expected in exact.items():
+    if md.version(package) != expected:
+        raise SystemExit(1)
+for package, minimum_version in minimum.items():
+    if version_tuple(md.version(package)) < version_tuple(minimum_version):
+        raise SystemExit(1)
+for package in present:
+    md.version(package)
+PY
+  then
+    echo "UI dependencies missing or stale; installing recorder dependencies"
+    install_ui_deps
+  fi
+fi
 # -----------------------------
 # Trainer server env
 # -----------------------------
 export DATA_DIR="${DATA_DIR}"
 export STATIC_DIR="${ROOTDIR}/static"
 export PERSONAL_DIR="${DATA_DIR}/personal_samples"
+export CAPTURED_DIR="${DATA_DIR}/captured_audio"
+export NEGATIVE_DIR="${DATA_DIR}/negative_samples"
+export TRAINED_WAKE_WORDS_DIR="${DATA_DIR}/trained_wake_words"
 
 # IMPORTANT: leave training venv creation to /api/train inside trainer_server.py
 # but still set TRAIN_CMD so the server knows how to invoke training once ready

train_wake_word

@@ -150,6 +150,7 @@ if ${CLEANUP_WORK_DIR} ; then
       "${DATA_DIR}/work/wake_word_samples" \
       "${DATA_DIR}/work/wake_word_samples_augmented" \
       "${DATA_DIR}/work/personal_augmented_features" \
+      "${DATA_DIR}/work/reviewed_negative_features" \
       "${DATA_DIR}/work/last_wake_word" || :
 fi