Release NVIDIA WakeWord Trainer v11

.github/workflows/docker-publish.yml

@@ -7,7 +7,7 @@ on:
   workflow_dispatch:
 
 permissions:
-  contents: read
+  contents: write
   packages: write
 
 concurrency:
@@ -41,6 +41,8 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          flavor: |
+            latest=false
           tags: |
             type=raw,value=latest
             type=ref,event=tag
@@ -56,3 +58,69 @@ jobs:
           labels: ${{ steps.meta.outputs.labels }}
           cache-from: type=gha,scope=mww-trainer-nvidia-docker
           cache-to: type=gha,mode=max,scope=mww-trainer-nvidia-docker
+
+      - name: Docker metadata (Blackwell)
+        id: meta-blackwell
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          flavor: |
+            latest=false
+          tags: |
+            type=raw,value=blackwell
+            type=ref,event=tag,suffix=-blackwell
+
+      - name: Build and push Blackwell image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: dockerfile.blackwell
+          platforms: linux/amd64
+          push: true
+          tags: ${{ steps.meta-blackwell.outputs.tags }}
+          labels: ${{ steps.meta-blackwell.outputs.labels }}
+          cache-from: type=gha,scope=mww-trainer-nvidia-docker-blackwell
+          cache-to: type=gha,mode=max,scope=mww-trainer-nvidia-docker-blackwell
+
+      - name: Create release notes
+        if: startsWith(github.ref, 'refs/tags/')
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          TAG_NAME: ${{ github.ref_name }}
+          REPO: ${{ github.repository }}
+        run: |
+          set -euo pipefail
+
+          title="microWakeWord Nvidia Trainer ${TAG_NAME}"
+          generated_notes="$(mktemp)"
+          release_notes="$(mktemp)"
+
+          gh api "repos/${REPO}/releases/generate-notes" \
+            -f tag_name="${TAG_NAME}" \
+            -f target_commitish="${GITHUB_SHA}" \
+            --jq '.body' > "${generated_notes}"
+
+          {
+            echo "## Docker Images"
+            echo
+            echo "- \`ghcr.io/tatertotterson/microwakeword:${TAG_NAME}\`"
+            echo "- \`ghcr.io/tatertotterson/microwakeword:latest\`"
+            echo "- \`ghcr.io/tatertotterson/microwakeword:${TAG_NAME}-blackwell\`"
+            echo "- \`ghcr.io/tatertotterson/microwakeword:blackwell\`"
+            echo
+            cat "${generated_notes}"
+          } > "${release_notes}"
+
+          if gh release view "${TAG_NAME}" >/dev/null 2>&1; then
+            gh release edit "${TAG_NAME}" \
+              --title "${title}" \
+              --notes-file "${release_notes}" \
+              --latest \
+              --verify-tag
+          else
+            gh release create "${TAG_NAME}" \
+              --title "${title}" \
+              --notes-file "${release_notes}" \
+              --latest \
+              --verify-tag
+          fi

README.md

@@ -7,7 +7,7 @@
   <a href="https://taterassistant.com">taterassistant.com</a>
 </h3>
 
-Train custom microWakeWord models in Docker with NVIDIA/CUDA acceleration, generated Piper samples, device-captured samples, reviewed false-wake negatives, live training logs, and prebuilt Tater firmware flashing.
+Train custom microWakeWord models in Docker with NVIDIA/CUDA acceleration, generated Piper samples, device-captured samples, reviewed false-wake negatives, live training logs, and local wake-word links for Tater Native satellites.
 
 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.
 
@@ -22,9 +22,22 @@ docker pull ghcr.io/tatertotterson/microwakeword:latest
 Tagged releases also publish matching immutable image tags:
 
 ```bash
-docker pull ghcr.io/tatertotterson/microwakeword:v5
+docker pull ghcr.io/tatertotterson/microwakeword:v11
 ```
 
+RTX 50-series / Blackwell GPUs use a separate image with CUDA 12.8 and a
+Python 3.13 TensorFlow build for `sm_120`:
+
+```bash
+docker pull ghcr.io/tatertotterson/microwakeword:blackwell
+docker pull ghcr.io/tatertotterson/microwakeword:v11-blackwell
+```
+
+Use the Blackwell image only for RTX 50-series cards. It includes the
+community-built TensorFlow wheel from
+[chivitiH/tensorflow-blackwell-python313](https://github.com/chivitiH/tensorflow-blackwell-python313),
+which is unofficial and licensed CC BY-NC 4.0.
+
 ---
 
 ## Run The Container
@@ -38,16 +51,19 @@ docker run -d \
   ghcr.io/tatertotterson/microwakeword:latest
 ```
 
-Use a version tag such as `ghcr.io/tatertotterson/microwakeword:v5` when you want to pin a known release instead of tracking `latest`.
+Use a version tag such as `ghcr.io/tatertotterson/microwakeword:v11` when you want to pin a known release instead of tracking `latest`.
+For RTX 50-series cards, use `ghcr.io/tatertotterson/microwakeword:blackwell`
+or a pinned tag such as `ghcr.io/tatertotterson/microwakeword:v11-blackwell`
+in the same `docker run` command.
 
 The flags:
 
 - `--gpus all` enables GPU acceleration.
-- `--network host` lets the container receive mDNS/zeroconf traffic for ESPHome auto-detect.
+- `--network host` exposes the trainer server directly so satellites can send captured audio and load trained wake-word files.
 - `-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.
+- `-v $(pwd):/data` persists models, downloaded voices, datasets, samples, and generated wake-word artifacts.
 
-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.
+If you do not use host networking, publish the trainer port and make sure satellites can reach it from your LAN.
 
 Open:
 
@@ -55,31 +71,37 @@ Open:
 http://localhost:8789
 ```
 
-If you change `REC_PORT`, open that port instead and use the same port in the ESPHome `Trainer App URL`.
+If you change `REC_PORT`, open that port instead and use the same port in the satellite `Trainer App URL`.
 
 ---
 
 ## What The UI Does
 
 - `Trainer` starts a wake-word session, shows positive/negative sample counts, and launches training.
-- `Captured Audio` reviews clips sent by ESPHome sats, including wake hits, close misses, and false wakes.
+- `Captured Audio` reviews clips sent by Tater Native or ESPHome sats, including wake hits, close misses, and false wakes.
 - `Samples` plays, removes, clears, and manually imports personal or negative samples.
-- `Firmware` pulls verified prebuilt Tater firmware images from GitHub and flashes supported satellites over OTA.
-- Popup consoles show colorized training and firmware logs while long-running jobs are active.
+- `Wake Words` lists locally trained JSON/model links for live wake-word switching in Tater.
+- Popup consoles show colorized training logs while long-running jobs are active.
 
 ---
 
 ## Captured Audio Workflow
 
-To collect samples from a sat, flash it with the Tater firmware from [TaterTotterson/microWakeWords](https://github.com/TaterTotterson/microWakeWords). The `Firmware` tab can pull verified prebuilt OTA images from that repo for fast firmware updates.
+To collect samples from a sat, point its trainer feedback setting at this app. Tater Native satellites use the native settings popup in Tater. Older ESPHome satellites can still use their device entities.
 
-After flashing, the device exposes ESPHome entities for capture setup:
+For Tater Native satellites, enable trainer feedback in Tater:
+
+- `Send Good Wakes To Trainer` toggles upload of confirmed wake-word triggers.
+- `Send Close Misses To Trainer` toggles upload of near misses.
+- `Trainer App URL` sets the trainer address, for example `http://trainer.local:8789` or `http://<trainer-ip>:8789`.
+
+For older ESPHome firmware, the equivalent capture setup is exposed as device entities:
 
 - `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:
+Satellites send raw captured audio to:
 
 ```text
 /api/upload_captured_audio_raw
@@ -153,6 +175,8 @@ Personal samples are optional. Training can run with zero personal samples after
 
 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.
 
+On RTX 50-series / Blackwell GPUs, the Blackwell Docker image keeps sample generation and augmentation in the normal Python 3.12 trainer environment, then runs only the TensorFlow training/export stage in `/data/.venv-blackwell` with Python 3.13 and the Blackwell-native TensorFlow wheel.
+
 ---
 
 ## Language Support
@@ -183,20 +207,16 @@ After those assets are prepared, later runs reuse the local copies unless the mo
 
 ---
 
-## Firmware Flashing
+## Trained Wake Words
 
-The `Firmware` tab flashes prebuilt Tater firmware for supported ESPHome satellites.
+The `Wake Words` tab lists locally trained wake-word packages from `/data/trained_wake_words/`.
 
-- Downloads the latest prebuilt firmware manifest plus OTA and USB factory images from `TaterTotterson/microWakeWords`.
-- Verifies downloaded images by size and SHA before upload.
-- 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 the selected OTA target for each firmware family.
-- Flashes the prebuilt factory image over Browser USB for first installs or recovery when opened in Chrome or Edge.
-- Lists locally trained wake words from `/data/trained_wake_words/` for live model switching.
-- Streams download, verification, and OTA upload progress in a colorized firmware console.
+- Copy the JSON URL into the Tater Native satellite settings to switch wake words live.
+- Open the JSON or model links directly for quick inspection.
+- The JSON includes the matching model path plus Tater tuning metadata.
+- No firmware flashing happens from this trainer app anymore.
 
-You usually only flash for firmware updates. New satellites, or devices older than Tater firmware `3.0.3`, need one USB flash first before OTA updates and live wake-word switching are available.
+Use the main Tater app for satellite firmware updates and USB flashing.
 
 ---
 
@@ -209,14 +229,50 @@ Successful runs produce timestamped training output folders such as:
 /data/output/<timestamp>-<wake_word>-<samples>-<steps>/<wake_word>.json
 ```
 
-The trainer also syncs firmware-ready artifacts into:
+The trainer also syncs Tater-ready wake-word 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.
+The `Wake Words` tab uses `/data/trained_wake_words/` to populate the local wake-word links.
+
+The JSON keeps the standard microWakeWord fields for compatibility:
+
+```json
+{
+  "micro": {
+    "probability_cutoff": 0.97,
+    "sliding_window_size": 5
+  }
+}
+```
+
+It also includes Tater Native metadata used by newer satellites and the Tater settings UI:
+
+```json
+{
+  "model_format": "tflite_stream_state_internal_quant",
+  "quantization": "int8",
+  "sample_rate": 16000,
+  "tater_native": {
+    "format_version": 1,
+    "wake_threshold": 0.97,
+    "wake_sliding_window": 5,
+    "close_miss_threshold": 0.78,
+    "frontend": {
+      "name": "tflm_microfrontend",
+      "sample_rate": 16000,
+      "feature_duration_ms": 30,
+      "feature_step_ms": 10,
+      "feature_size": 40
+    }
+  }
+}
+```
+
+Calibration metrics are included under `calibration` so false accepts/hour and recall can be surfaced in the UI.
 
 ---
 
@@ -233,7 +289,6 @@ That removes:
 - cached datasets
 - training environments
 - trained models
-- downloaded firmware images
 
 ---
 
@@ -243,7 +298,7 @@ That removes:
 - Negative samples are optional but useful for reducing false wakes.
 - The UI server is `trainer_server.py`.
 - The launcher is `run.sh`.
-- Firmware capture settings live on the ESPHome device and can be toggled from the device entities after flashing.
+- Trainer capture settings live in Tater for Tater Native satellites, and on device entities for older ESPHome satellites.
 
 ---
 
@@ -253,3 +308,4 @@ Built on top of:
 
 - [microWakeWord](https://github.com/kahrendt/microWakeWord)
 - [piper-sample-generator](https://github.com/rhasspy/piper-sample-generator)
+- [tensorflow-blackwell-python313](https://github.com/chivitiH/tensorflow-blackwell-python313) for the optional RTX 50-series / Blackwell image

cli/calibrate_detector.py

@@ -18,14 +18,14 @@ from microwakeword.data import FeatureHandler
 from microwakeword.inference import Model
 
 
-DEFAULT_WINDOW_SIZES = [3, 4, 5, 6, 7]
-DEFAULT_TARGET_FAPH = float(os.environ.get("MWW_CALIBRATION_TARGET_FAPH", "1.0"))
+DEFAULT_WINDOW_SIZES = [4, 5, 6, 7]
+DEFAULT_TARGET_FAPH = float(os.environ.get("MWW_CALIBRATION_TARGET_FAPH", "0.25"))
 DEFAULT_COOLDOWN_SLICES = int(os.environ.get("MWW_CALIBRATION_COOLDOWN_SLICES", "25"))
 DEFAULT_POSITIVE_SKIP_SLICES = int(
     os.environ.get("MWW_CALIBRATION_POSITIVE_SKIP_SLICES", "25")
 )
 DEFAULT_CUTOFF_STEP = float(os.environ.get("MWW_CALIBRATION_CUTOFF_STEP", "0.01"))
-DEFAULT_CUTOFF_MIN = float(os.environ.get("MWW_CALIBRATION_CUTOFF_MIN", "0.00"))
+DEFAULT_CUTOFF_MIN = float(os.environ.get("MWW_CALIBRATION_CUTOFF_MIN", "0.85"))
 DEFAULT_CUTOFF_MAX = float(os.environ.get("MWW_CALIBRATION_CUTOFF_MAX", "1.00"))
 
 

cli/setup_blackwell_venv

@@ -0,0 +1,112 @@
+#!/bin/bash
+set -euo pipefail
+
+PROGDIR="$(dirname "$(realpath "$0")")"
+ROOTDIR="$(dirname "${PROGDIR}")"
+
+KNOWN_ARGS=( data-dir force python )
+source "${PROGDIR}/shell.functions"
+
+if [ ${#UNKNOWN_ARGS[@]} -gt 0 ] ; then
+    echo "Unknown argument(s): ${UNKNOWN_ARGS[*]}" >&2
+    HELP=true
+fi
+
+if [ "${HELP}" == "true" ] ; then
+    cat <<EOF >&2
+Usage: setup_blackwell_venv [ --data-dir=/data ] [ --force ] [ --python=python3.13 ]
+
+Creates /data/.venv-blackwell for RTX 50 / Blackwell TensorFlow training.
+Sample generation and augmentation continue to use /data/.venv.
+
+Environment overrides:
+MWW_BLACKWELL_TF_WHEEL_URL: TensorFlow Blackwell wheel URL.
+
+EOF
+    exit 1
+fi
+
+[ -n "${DATA_DIR}" ] && DATA_DIR="$(realpath "${DATA_DIR}")"
+[ -d "${DATA_DIR}" ] || {
+    echo "Data directory '${DATA_DIR}' doesn't exist." >&2
+    exit 1
+}
+
+PYTHON="${PYTHON:-python3.13}"
+VENV="${DATA_DIR}/.venv-blackwell"
+MARKER="${VENV}/.mww-blackwell-venv"
+TF_WHEEL_URL="${MWW_BLACKWELL_TF_WHEEL_URL:-https://github.com/chivitiH/tensorflow-blackwell-python313/releases/download/v2.22.0-selfbuilt/tensorflow-2.22.0.dev0+selfbuilt-cp313-cp313-linux_x86_64.whl}"
+
+if ! command -v "${PYTHON}" >/dev/null 2>&1 ; then
+    echo "Python 3.13 is required for the Blackwell TensorFlow wheel. Missing: ${PYTHON}" >&2
+    exit 1
+fi
+
+if [ "${FORCE:-false}" != "true" ] && [ -x "${VENV}/bin/python" ] && [ -f "${MARKER}" ] ; then
+    echo "   Blackwell TensorFlow venv found (skipping setup_blackwell_venv)"
+    exit 0
+fi
+
+echo "===== Setting up Blackwell TensorFlow environment ${VENV} ====="
+rm -rf "${VENV}" || :
+"${PYTHON}" -m venv --upgrade-deps "${VENV}"
+source "${VENV}/bin/activate"
+
+export PIP_PROGRESS_BAR=off
+export PIP_NO_COLOR=1
+export PIP_QUIET=0
+
+pip_install() {
+    if $VERBOSE ; then
+        pip install "$@" || return 1
+    else
+        { pip install "$@" || return 1 ; } | stdbuf -i0 -o0 tr -d '[:print:]' | stdbuf -i0 -o0 tr '\n' '.'
+    fi
+    echo
+}
+
+echo "   ===== Installing Blackwell TensorFlow wheel ====="
+pip_install --upgrade pip setuptools wheel
+pip_install "${TF_WHEEL_URL}"
+
+echo "   ===== Installing microWakeWord training dependencies ====="
+pip_install \
+    audiomentations \
+    audio_metadata \
+    datasets \
+    mmap_ninja \
+    pymicro-features \
+    pyyaml \
+    webrtcvad-wheels \
+    ai-edge-litert \
+    numpy-minmax \
+    numpy-rms \
+    absl-py \
+    "numpy==2.3.5"
+
+echo "   ===== Checking microwakeword ====="
+MWW="${DATA_DIR}/tools/microWakeWord"
+if [ ! -d "${MWW}" ] || [ -n "$(git -C "${MWW}" status --porcelain 2>/dev/null || true)" ] ; then
+    rm -rf "${MWW}" || :
+    mkdir -p "${DATA_DIR}/tools"
+    echo "   Cloning micro-wake-word to ${DATA_DIR}/tools"
+    git clone https://github.com/TaterTotterson/micro-wake-word "${MWW}" &>/dev/null
+fi
+echo "   Installing microwakeword into Blackwell venv"
+pip_install --no-deps -e "${MWW}"
+
+echo "   ===== Testing Blackwell TensorFlow environment ====="
+"${VENV}/bin/python" - <<'PY'
+import tensorflow as tf
+from ai_edge_litert.interpreter import Interpreter
+from microwakeword.data import FeatureHandler
+from microwakeword.inference import Model
+
+print("TensorFlow:", tf.__version__)
+print("CUDA build:", tf.test.is_built_with_cuda())
+print("GPU:", tf.config.list_physical_devices("GPU"))
+print("microWakeWord Blackwell imports available")
+PY
+
+touch "${MARKER}"
+echo "Blackwell TensorFlow environment ready: ${VENV}"

cli/setup_mit_audio

@@ -25,9 +25,9 @@ fi
 mkdir -p "${DATA_DIR}/training_datasets/downloads" || :
 cd "${DATA_DIR}/training_datasets"
 
-AUDIO_URL="https://mcdermottlab.mit.edu/Reverb/IRMAudio/Audio.zip"
-AUDIO_ZIPFILE="MIT_RIR_Audio.zip"
-AUDIO_ZIP="./downloads/${AUDIO_ZIPFILE}"
+HF_RIR_REPO_ID="TaterTotterson/MIT_environmental_impulse_responses"
+HF_RIR_API_URL="https://huggingface.co/api/datasets/${HF_RIR_REPO_ID}"
+HF_RIR_SOURCE_KEY="hf_mit_environmental_impulse_responses"
 AUDIO_DIR="./mit_rirs"
 mkdir -p "${AUDIO_DIR}" || :
 AUDIO16K_DIR="./mit_rirs_16k"
@@ -35,10 +35,92 @@ mkdir -p "${AUDIO16K_DIR}" || :
 AUDIO_FILECOUNT="./downloads/mit_rir_filecount"
 AUDIO_IN_GLOB="*.wav"
 
-declare -A filecounts=( [${AUDIO_ZIPFILE}]=0 )
+declare -A filecounts=( [${HF_RIR_SOURCE_KEY}]=0 )
 get_filecounts filecounts "${AUDIO_FILECOUNT}"
 
-echo "===== Checking MIT_RIR ====="
+echo "===== Checking MIT environmental RIRs ====="
+
+download_hf_mit_rirs() {
+    source ${DATA_DIR}/.venv/bin/activate
+    python - "${HF_RIR_REPO_ID}" "${HF_RIR_API_URL}" "${AUDIO_DIR}" <<-'EOF'
+import json
+import sys
+import time
+import urllib.parse
+import urllib.request
+from pathlib import Path
+
+repo_id = sys.argv[1]
+api_url = sys.argv[2]
+audio_dir = Path(sys.argv[3])
+audio_dir.mkdir(parents=True, exist_ok=True)
+
+request = urllib.request.Request(api_url, headers={"User-Agent": "WakeWordTrainer/1.0"})
+with urllib.request.urlopen(request, timeout=30) as response:
+    metadata = json.loads(response.read().decode("utf-8"))
+
+files = sorted(
+    sibling.get("rfilename", "")
+    for sibling in metadata.get("siblings", [])
+    if str(sibling.get("rfilename", "")).startswith("16khz/")
+    and str(sibling.get("rfilename", "")).lower().endswith(".wav")
+)
+if not files:
+    raise SystemExit("Hugging Face MIT RIR dataset did not list any 16khz WAV files")
+
+print(f"   Found {len(files)} MIT environmental RIR files on Hugging Face mirror", flush=True)
+downloaded = 0
+skipped = 0
+
+def download_file(url: str, target: Path, rel: str):
+    tmp = target.with_suffix(target.suffix + ".incomplete")
+    for attempt in range(1, 4):
+        try:
+            if tmp.exists():
+                tmp.unlink()
+            with urllib.request.urlopen(url, timeout=30) as response:
+                with tmp.open("wb") as out:
+                    while True:
+                        chunk = response.read(1024 * 64)
+                        if not chunk:
+                            break
+                        out.write(chunk)
+            if not tmp.exists() or tmp.stat().st_size == 0:
+                raise RuntimeError("empty download")
+            tmp.replace(target)
+            return
+        except Exception as exc:
+            if tmp.exists():
+                tmp.unlink()
+            if attempt == 3:
+                raise RuntimeError(f"download failed for {rel}: {exc}") from exc
+            print(f"   Retry {attempt}/2 for {rel}: {exc}", flush=True)
+            time.sleep(2 * attempt)
+
+total = len(files)
+for idx, rel in enumerate(files, start=1):
+    target = audio_dir / rel
+    if target.exists() and target.stat().st_size > 0:
+        skipped += 1
+        if idx == 1 or idx % 25 == 0 or idx == total:
+            print(f"   MIT RIR download progress: {idx}/{total} files ({downloaded} downloaded, {skipped} reused)", flush=True)
+        continue
+    target.parent.mkdir(parents=True, exist_ok=True)
+    encoded = urllib.parse.quote(rel, safe="/")
+    url = f"https://huggingface.co/datasets/{repo_id}/resolve/main/{encoded}"
+    if idx == 1 or idx % 25 == 0 or idx == total:
+        print(f"   Downloading MIT RIR {idx}/{total}: {rel}", flush=True)
+    download_file(url, target, rel)
+    if not target.exists() or target.stat().st_size == 0:
+        raise SystemExit(f"download failed for {rel}")
+    downloaded += 1
+    if idx == 1 or idx % 25 == 0 or idx == total:
+        print(f"   MIT RIR download progress: {idx}/{total} files ({downloaded} downloaded, {skipped} reused)", flush=True)
+
+print(f"   Hugging Face MIT environmental RIR download complete ({downloaded} downloaded, {skipped} reused)", flush=True)
+print(f"   MIT environmental RIR files available: {len(files)}", flush=True)
+EOF
+}
 
 converter() {
     source ${DATA_DIR}/.venv/bin/activate
@@ -58,9 +140,9 @@ rir_out = Path(sys.argv[2])
 
 waves = list(rir_in.rglob("*.wav"))
 try:
-    print("   MIT RIR normalizing to 16k…")
+    print("   MIT environmental RIR normalizing to 16k…")
     # Normalize to 16k mono
-    for p in tqdm(waves, desc="   MIT_RIR (resample 16k mono)"):
+    for p in tqdm(waves, desc="   MIT environmental RIR (resample 16k mono)"):
         outfile = Path(rir_out / p.name)
         if outfile.exists():
             continue
@@ -70,14 +152,14 @@ try:
         if sr != 16000:
             a, _ = librosa.load(p, sr=16000, mono=True)
         write_wav(outfile, a, 16000)
-    print("   MIT RIR normalization complete")
+    print("   MIT environmental RIR normalization complete")
 except Exception as e2:
-    print(f"   MIT RIR fallback failed: {e2}")
+    print(f"   MIT environmental RIR preparation failed: {e2}")
     raise
 EOF
 }
 
-expected_filecount=${filecounts[${AUDIO_ZIPFILE}]}
+expected_filecount=${filecounts[${HF_RIR_SOURCE_KEY}]}
 actual_filecount=$(find "${AUDIO16K_DIR}" -name '*.wav' 2>/dev/null | wc -l) || :
 write_filecount=false
 
@@ -85,24 +167,16 @@ if [ "${actual_filecount}" -ne 0 ] && [ "${actual_filecount}" -eq "${expected_fi
     echo "   Existing ${AUDIO16K_DIR} valid"
 else
     actual_filecount=$(find "${AUDIO_DIR}" -name "${AUDIO_IN_GLOB}" 2>/dev/null | wc -l) || :
-    if [ "${actual_filecount}" -eq 0 ] || [ "${actual_filecount}" -ne "${expected_filecount}" ] ; then
-        if [ ! -f "${AUDIO_ZIP}" ]  ; then
-            echo "   Downloading ${AUDIO_ZIPFILE}"
-            curl -sfL "${AUDIO_URL}" -o "${AUDIO_ZIP}"
-        fi
-
+    if [ "${actual_filecount}" -eq 0 ] || [ "${expected_filecount}" -eq 0 ] || [ "${actual_filecount}" -ne "${expected_filecount}" ] ; then
         rm -rf "${AUDIO_DIR}" || :
-        echo "   Unzipping ${AUDIO_ZIPFILE}"
-        unzip -u -q -d "${AUDIO_DIR}" "${AUDIO_ZIP}"
-    fi
-    if "${CLEANUP_ARCHIVES}" && [ -f "${AUDIO_ZIP}" ] ; then
-        echo "   Cleaning up ${AUDIO_ZIPFILE}"
-        rm -rf "${AUDIO_ZIP}"
+        mkdir -p "${AUDIO_DIR}" || :
+        echo "   Downloading MIT environmental impulse responses from Hugging Face mirror"
+        download_hf_mit_rirs
     fi
 
     converter
     actual_filecount=$(find "${AUDIO16K_DIR}" -name "*.wav" 2>/dev/null | wc -l) || :
-    filecounts[${AUDIO_ZIPFILE}]="${actual_filecount}"
+    filecounts[${HF_RIR_SOURCE_KEY}]="${actual_filecount}"
     write_filecount=true
 fi
 
@@ -110,15 +184,10 @@ if ${write_filecount} ; then
     write_filecounts filecounts "${AUDIO_FILECOUNT}"
 fi
 
-if "${CLEANUP_ARCHIVES}" && [ -f "${AUDIO_ZIP}" ] ; then
-    echo "   Cleaning up ${AUDIO_ZIPFILE}"
-    rm -rf "${AUDIO_ZIP}"
-fi
-
 if "${CLEANUP_INTERMEDIATE_FILES}" && [ -d "${AUDIO_DIR}" ]; then
     echo "   Cleaning up ${AUDIO_DIR}"
     rm -rf "${AUDIO_DIR}"
 fi
 
-echo "   MIT_RIR complete"
+echo "   MIT environmental RIRs complete"
 exit 0

cli/wake_word_sample_trainer

@@ -84,6 +84,21 @@ if [ "${IS_BLACKWELL}" = "true" ]; then
   echo "ℹ️  Using GPU compatibility retries; CPU fallback is ${ALLOW_CPU_FALLBACK} (override with MWW_ALLOW_CPU_FALLBACK=true|false)."
 fi
 
+BLACKWELL_TF_MODE="${MWW_BLACKWELL_TF:-auto}"
+BLACKWELL_TF_REQUIRED="false"
+BLACKWELL_TF_ACTIVE="false"
+case "${BLACKWELL_TF_MODE,,}" in
+  1|true|yes|on|required)
+    BLACKWELL_TF_REQUIRED="true"
+    ;;
+  0|false|no|off|disabled)
+    BLACKWELL_TF_MODE="disabled"
+    ;;
+  *)
+    BLACKWELL_TF_MODE="auto"
+    ;;
+esac
+
 # Enable driver-side PTX JIT fallback when ptxas/nvlink are unavailable.
 if [ -z "${XLA_FLAGS:-}" ]; then
   export XLA_FLAGS="--xla_gpu_unsafe_fallback_to_driver_on_ptxas_not_found"
@@ -238,6 +253,32 @@ fi
 echo "   Wrote training_parameters.yaml"
 rm -rf "${WORK_DIR}/trained_models/wakeword"
 
+if [ "${IS_BLACKWELL}" = "true" ] && [ "${BLACKWELL_TF_MODE}" != "disabled" ]; then
+  BLACKWELL_SETUP="${PROGDIR}/setup_blackwell_venv"
+  BLACKWELL_PYTHON="${DATA_DIR}/.venv-blackwell/bin/python"
+
+  if [ -x "${BLACKWELL_SETUP}" ] && command -v python3.13 >/dev/null 2>&1; then
+    echo "↪️  Preparing Blackwell-native TensorFlow training environment."
+    if "${BLACKWELL_SETUP}" --data-dir="${DATA_DIR}"; then
+      PYTHON_BIN="${BLACKWELL_PYTHON}"
+      BLACKWELL_TF_ACTIVE="true"
+      echo "✅ Blackwell TensorFlow training enabled: ${PYTHON_BIN}"
+    else
+      if [ "${BLACKWELL_TF_REQUIRED}" = "true" ]; then
+        echo "❌ Blackwell TensorFlow setup failed and MWW_BLACKWELL_TF is required." >&2
+        exit 1
+      fi
+      echo "⚠️  Blackwell TensorFlow setup failed; continuing with compatibility retries."
+    fi
+  else
+    if [ "${BLACKWELL_TF_REQUIRED}" = "true" ]; then
+      echo "❌ Blackwell TensorFlow was required, but python3.13/setup_blackwell_venv is unavailable." >&2
+      exit 1
+    fi
+    echo "ℹ️  Blackwell TensorFlow image support not available; continuing with compatibility retries."
+  fi
+fi
+
 wake_word_filename="$(
   echo "${WAKE_WORD}" \
     | tr '[:upper:]' '[:lower:]' \
@@ -261,11 +302,11 @@ TRAIN_ARGS=(
   --test_tflite_streaming_quantized 1
   --use_weights best_weights
   mixednet
-  --pointwise_filters "64,64,64,64"
+  --pointwise_filters "128,128,128,128"
   --repeat_in_block "1,1,1,1"
   --mixconv_kernel_sizes "[5], [7,11], [9,15], [23]"
   --residual_connection "0,0,0,0"
-  --first_conv_filters 32
+  --first_conv_filters 64
   --first_conv_kernel_size 5
   --stride 2
 )
@@ -345,6 +386,7 @@ fi
 TRAINING_DONE="false"
 
 echo "🏋️ Starting model training and TFLite export (this is the longest stage)…"
+echo "🧠 Model quality: high_accuracy_plus"
 if run_attempt "Attempt 1/3: GPU training (default runtime profile)" ; then
   echo "✅ Training complete (GPU path)."
   TRAINING_DONE="true"
@@ -454,8 +496,10 @@ from pathlib import Path
 json_path = Path(os.environ["JSON_PATH"])
 calibration_path = Path(os.environ.get("CALIBRATION_PATH", ""))
 language = (os.environ.get("LANGUAGE", "en") or "en").strip().lower()
-probability_cutoff = 0.97
-sliding_window_size = 5
+probability_cutoff = 0.85
+sliding_window_size = 4
+strict_min_close_miss_threshold = 0.68
+calibration = {}
 
 if calibration_path.exists():
     try:
@@ -469,21 +513,63 @@ if calibration_path.exists():
     except Exception as exc:
         print(f"⚠️ Failed to read detector calibration ({exc}); using defaults.")
 
+probability_cutoff = round(probability_cutoff, 3)
+sliding_window_size = max(1, min(10, int(sliding_window_size)))
+selected_metrics = calibration.get("selected_metrics") if isinstance(calibration.get("selected_metrics"), dict) else {}
+evaluation = calibration.get("evaluation") if isinstance(calibration.get("evaluation"), dict) else {}
+close_miss_threshold = max(
+    0.01,
+    min(0.99, round(max(strict_min_close_miss_threshold, probability_cutoff - 0.17), 3)),
+)
+
 meta = {
     "type": "micro",
     "wake_word": os.environ["WAKE_WORD_TITLE"],
+    "label": os.environ["WAKE_WORD_TITLE"].replace("_", " ").title(),
     "author": "Tater Totterson",
     "website": "https://github.com/TaterTotterson/microWakeWord-Trainer-Nvidia-Docker.git",
     "model": os.environ["TFLITE_FILENAME"],
     "trained_languages": [language],
     "version": 2,
+    "model_format": "tflite_stream_state_internal_quant",
+    "quantization": "int8",
+    "sample_rate": 16000,
     "micro": {
-        "probability_cutoff": round(probability_cutoff, 2),
+        "probability_cutoff": probability_cutoff,
         "sliding_window_size": sliding_window_size,
         "feature_step_size": 10,
         "tensor_arena_size": 30000,
         "minimum_esphome_version": "2024.7.0",
     },
+    "tater_native": {
+        "format_version": 1,
+        "wake_threshold": probability_cutoff,
+        "wake_sliding_window": sliding_window_size,
+        "close_miss_threshold": close_miss_threshold,
+        "frontend": {
+            "name": "tflm_microfrontend",
+            "sample_rate": 16000,
+            "feature_duration_ms": 30,
+            "feature_step_ms": 10,
+            "feature_size": 40,
+            "input_feature_frames": 2,
+            "lower_band_limit": 125.0,
+            "upper_band_limit": 7500.0,
+        },
+        "recommended_for": ["tater-native-satellite", "voice-pe"],
+    },
+    "calibration": {
+        "target_false_accepts_per_hour": calibration.get("target_false_accepts_per_hour"),
+        "selected_false_accepts_per_hour_limit": calibration.get("selected_false_accepts_per_hour_limit"),
+        "recall": selected_metrics.get("recall"),
+        "false_accepts_per_hour": selected_metrics.get("false_accepts_per_hour"),
+        "ambient_hours": selected_metrics.get("ambient_hours"),
+        "positive_dataset": evaluation.get("positive_dataset"),
+        "ambient_dataset": evaluation.get("ambient_dataset"),
+        "positive_tracks": evaluation.get("positive_tracks"),
+        "ambient_tracks": evaluation.get("ambient_tracks"),
+        "generated_at": calibration.get("generated_at"),
+    },
 }
 json_path.write_text(json.dumps(meta, indent=4) + "\n", encoding="utf-8")
 PY

dockerfile.blackwell

@@ -0,0 +1,54 @@
+# RTX 50 / Blackwell image
+FROM nvidia/cuda:12.8.1-cudnn-devel-ubuntu24.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV CUDA_HOME=/usr/local/cuda
+ENV PATH=/usr/local/cuda/bin:${PATH}
+ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
+ENV MWW_BLACKWELL_IMAGE=1
+ENV MWW_BLACKWELL_TF=auto
+ENV MWW_BLACKWELL_TF_WHEEL_URL=https://github.com/chivitiH/tensorflow-blackwell-python313/releases/download/v2.22.0-selfbuilt/tensorflow-2.22.0.dev0+selfbuilt-cp313-cp313-linux_x86_64.whl
+
+# System deps. Python 3.12 remains the main trainer/runtime venv, while
+# Python 3.13 is used only for the Blackwell TensorFlow training step.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    software-properties-common ca-certificates curl git wget unzip patch \
+    ninja-build nano less \
+ && add-apt-repository -y ppa:deadsnakes/ppa \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+    python3.12 python3.12-venv python3.12-dev \
+    python3.13 python3.13-venv python3.13-dev \
+    python3-pip python-is-python3 \
+ && ldconfig \
+ && rm -rf /var/lib/apt/lists/* \
+ && mkdir -p /data
+
+# Trainer UI port
+EXPOSE 8789
+
+# Script root
+WORKDIR /root/mww-scripts
+
+# Bash environment
+COPY --chown=root:root --chmod=0755 .bashrc /root/
+
+# Root-level entrypoints
+COPY --chown=root:root --chmod=0755 \
+    train_wake_word \
+    run.sh \
+    trainer_server.py \
+    requirements.txt \
+    /root/mww-scripts/
+
+# CLI folder
+COPY --chown=root:root cli/ /root/mww-scripts/cli/
+
+# Make all CLI scripts executable (avoids "Permission denied")
+RUN chmod -R a+x /root/mww-scripts/cli
+
+# Static UI for trainer
+COPY --chown=root:root --chmod=0644 static/index.html /root/mww-scripts/static/index.html
+
+# trainer server
+CMD ["/bin/bash", "-lc", "/root/mww-scripts/run.sh"]

run.sh

@@ -30,7 +30,6 @@ install_ui_deps() {
     "fastapi==${FASTAPI_VERSION}" \
     "uvicorn[standard]==${UVICORN_VERSION}" \
     "python-multipart==${PY_MULTIPART_VERSION}" \
-    "zeroconf>=0.132.2" \
     "silero-vad>=5.0.0" \
     "numpy>=1.24.0"
 }
@@ -79,7 +78,6 @@ exact = {
 minimum = {
     "silero-vad": "5.0.0",
     "numpy": "1.24.0",
-    "zeroconf": "0.132.2",
 }
 present = ("torch",)