mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-31 19:16:29 +00:00
fix(wake): run openWakeWord on tflite on macOS ARM64
openWakeWord's ONNX backend returns near-zero scores on Apple Silicon (dscripka/openWakeWord#336), so "Hey Hermes" never crossed the 0.5 threshold: the listener armed, the microphone worked, and nothing fired. Bisecting the pipeline puts the fault in exactly one stage — feeding the same audio through both backends, the melspectrogram front-end is bit-identical (maxdiff 0.00000) and the wake classifier agrees on identical features, while the shared embedding model diverges by 45.44. Cross-feeding confirms it: tflite features scored through the *onnx* classifier give 0.9948 vs 0.000009 for onnx features. A telling secondary symptom is that scores fall as input gets louder (0.5x -> 0.00031, 8x -> 0.000066), which is garbage inference rather than a weak detection. Selecting tflite in config alone does not fix it. openWakeWord hardcodes `import tflite_runtime.interpreter` but declares tflite-runtime for `platform_system == "Linux"` only; on macOS the equivalent wheel is ai-edge-litert, so that import always fails and model.py silently downgrades back to onnx. The result is a detector that reports itself listening and can never fire. - default the backend per platform (tflite on macOS ARM64, onnx elsewhere) instead of hardcoding onnx, and pick the matching bundled model artifact - bridge tflite_runtime -> ai_edge_litert through sys.modules, in-process, with no writes to site-packages - refuse the silent onnx downgrade on macOS ARM64 and report the missing runtime through check_wake_word_requirements() so the GUI surfaces an actionable hint rather than arming a dead ear - lazy-install ai-edge-litert via its own feature key, because lazy-dep specs cannot carry PEP 508 markers (_spec_is_safe rejects ";") An explicit `inference_framework` in config still wins, so anyone pinning a backend keeps it. Verified on macOS 26.5.2 / M-series: "hey hermes" scores 0.0005 on onnx and 0.9423 on tflite from the same clip, with cross-phrase controls at 0.0003. Live over-the-air through the real microphone fires 4/4 utterances (peak 0.9532).
This commit is contained in:
Ben Barclay
parent
4dd4708346
commit
76ac204846
7 changed files with 222 additions and 8 deletions
|
|
@ -2371,7 +2371,11 @@ DEFAULT_CONFIG = {
|
|||
# ...) OR a path to a custom .onnx/.tflite model for another phrase.
|
||||
# See the wake-word docs for the custom-model training guide.
|
||||
"model": "hey_hermes",
|
||||
"inference_framework": "onnx", # "onnx" | "tflite"
|
||||
# "" (auto — tflite on macOS ARM64, onnx elsewhere) | "onnx" | "tflite".
|
||||
# openWakeWord's onnx backend scores near-zero on macOS ARM64
|
||||
# (dscripka/openWakeWord#336), so auto avoids a listener that arms
|
||||
# but never fires. Set explicitly only to override that choice.
|
||||
"inference_framework": "",
|
||||
},
|
||||
"sherpa": {
|
||||
# Optional path to a sherpa-onnx KWS model directory. Empty =
|
||||
|
|
|
|||
|
|
@ -188,6 +188,11 @@ wake = [
|
|||
"pvporcupine==4.0.3",
|
||||
"sounddevice==0.5.5",
|
||||
"numpy==2.4.3",
|
||||
# openWakeWord's onnx embedding model scores near-zero on macOS ARM64
|
||||
# (dscripka/openWakeWord#336), so the wake word runs on tflite there.
|
||||
# Upstream declares tflite-runtime for Linux only; ai-edge-litert is the
|
||||
# macOS equivalent, bridged in tools/wake_word.py.
|
||||
"ai-edge-litert==2.1.6; platform_system == 'Darwin'",
|
||||
]
|
||||
honcho = ["honcho-ai==2.2.0"]
|
||||
# Cloud memory providers — opt-in, lazy-installed via tools/lazy_deps.py
|
||||
|
|
|
|||
|
|
@ -224,11 +224,13 @@ def test_bundled_hey_hermes_model_ships_on_disk():
|
|||
def test_openwakeword_default_resolves_to_bundled_model(monkeypatch, model_value):
|
||||
# The default (and any "hey_hermes" alias) must load the bundled file, not be
|
||||
# passed through as a bogus built-in name that openWakeWord can't resolve.
|
||||
# Which artifact is bundled follows the platform's default backend (tflite on
|
||||
# macOS ARM64, where openWakeWord's onnx path scores near-zero).
|
||||
calls = _install_fake_openwakeword(monkeypatch)
|
||||
sub = {} if model_value is None else {"model": model_value}
|
||||
ww._OpenWakeWordEngine({"provider": "openwakeword", "openwakeword": sub})
|
||||
(downloaded,) = calls["download"]
|
||||
assert downloaded == [ww._bundled_wakeword_path("onnx")]
|
||||
assert downloaded == [ww._bundled_wakeword_path(ww.default_inference_framework())]
|
||||
|
||||
|
||||
def test_openwakeword_bundled_model_matches_framework(monkeypatch):
|
||||
|
|
@ -238,7 +240,78 @@ def test_openwakeword_bundled_model_matches_framework(monkeypatch):
|
|||
)
|
||||
(downloaded,) = calls["download"]
|
||||
assert downloaded == [ww._bundled_wakeword_path("tflite")]
|
||||
assert downloaded[0].endswith(".tflite")
|
||||
|
||||
|
||||
# ── platform-aware backend selection (openWakeWord onnx is broken on macOS ARM64,
|
||||
# upstream dscripka/openWakeWord#336) ────────────────────────────────────────
|
||||
|
||||
def test_default_framework_is_tflite_on_macos_arm64(monkeypatch):
|
||||
monkeypatch.setattr(ww.sys, "platform", "darwin")
|
||||
monkeypatch.setattr("platform.machine", lambda: "arm64")
|
||||
assert ww.default_inference_framework() == "tflite"
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"plat,machine",
|
||||
[("linux", "x86_64"), ("linux", "aarch64"), ("win32", "AMD64"), ("darwin", "x86_64")],
|
||||
)
|
||||
def test_default_framework_is_onnx_elsewhere(monkeypatch, plat, machine):
|
||||
# Only the broken platform changes behaviour; everyone else keeps onnx.
|
||||
monkeypatch.setattr(ww.sys, "platform", plat)
|
||||
monkeypatch.setattr("platform.machine", lambda: machine)
|
||||
assert ww.default_inference_framework() == "onnx"
|
||||
|
||||
|
||||
def test_explicit_framework_overrides_platform_default(monkeypatch):
|
||||
# An operator who pins a backend keeps it, even on macOS ARM64.
|
||||
calls = _install_fake_openwakeword(monkeypatch)
|
||||
monkeypatch.setattr(ww.sys, "platform", "darwin")
|
||||
monkeypatch.setattr("platform.machine", lambda: "arm64")
|
||||
ww._OpenWakeWordEngine(
|
||||
{"provider": "openwakeword", "openwakeword": {"inference_framework": "onnx"}}
|
||||
)
|
||||
(downloaded,) = calls["download"]
|
||||
assert downloaded == [ww._bundled_wakeword_path("onnx")]
|
||||
|
||||
|
||||
def test_macos_tflite_refuses_silent_onnx_downgrade(monkeypatch):
|
||||
# openWakeWord silently falls back to onnx when no tflite runtime imports.
|
||||
# On macOS ARM64 that lands on the broken backend, so we must raise instead
|
||||
# of arming a listener that can never fire.
|
||||
_install_fake_openwakeword(monkeypatch)
|
||||
monkeypatch.setattr(ww.sys, "platform", "darwin")
|
||||
monkeypatch.setattr("platform.machine", lambda: "arm64")
|
||||
monkeypatch.setattr(ww, "ensure_tflite_runtime", lambda: False)
|
||||
with pytest.raises(RuntimeError, match="ai-edge-litert"):
|
||||
ww._OpenWakeWordEngine({"provider": "openwakeword", "openwakeword": {}})
|
||||
|
||||
|
||||
def test_non_macos_tflite_falls_back_to_onnx(monkeypatch):
|
||||
# Off macOS the onnx backend works, so a missing tflite runtime is a
|
||||
# downgrade, not a hard failure.
|
||||
calls = _install_fake_openwakeword(monkeypatch)
|
||||
monkeypatch.setattr(ww.sys, "platform", "linux")
|
||||
monkeypatch.setattr("platform.machine", lambda: "x86_64")
|
||||
monkeypatch.setattr(ww, "ensure_tflite_runtime", lambda: False)
|
||||
ww._OpenWakeWordEngine(
|
||||
{"provider": "openwakeword", "openwakeword": {"inference_framework": "tflite"}}
|
||||
)
|
||||
(downloaded,) = calls["download"]
|
||||
assert downloaded == [ww._bundled_wakeword_path("onnx")]
|
||||
|
||||
|
||||
def test_requirements_report_missing_tflite_runtime(monkeypatch):
|
||||
# A missing runtime must surface as unavailable + an actionable hint rather
|
||||
# than an armed-but-deaf detector.
|
||||
monkeypatch.setattr(ww.sys, "platform", "darwin")
|
||||
monkeypatch.setattr("platform.machine", lambda: "arm64")
|
||||
monkeypatch.setattr(ww, "ensure_tflite_runtime", lambda: False)
|
||||
monkeypatch.setattr("tools.lazy_deps.is_available", lambda feature: feature != "wake.openwakeword.tflite")
|
||||
monkeypatch.setattr("tools.lazy_deps._allow_lazy_installs", lambda: False)
|
||||
monkeypatch.setattr(ww, "_audio_available", lambda: True)
|
||||
reqs = ww.check_wake_word_requirements({"provider": "openwakeword", "openwakeword": {}})
|
||||
assert reqs["available"] is False
|
||||
assert "ai-edge-litert" in reqs["hint"]
|
||||
|
||||
|
||||
# ── sherpa-onnx open-vocabulary engine ───────────────────────────────────
|
||||
|
|
|
|||
|
|
@ -140,6 +140,16 @@ LAZY_DEPS: dict[str, tuple[str, ...]] = {
|
|||
# ─── Wake word ("Hey Hermes") engines ──────────────────────────────────
|
||||
# Keep in sync with the `wake` extra in pyproject.toml. openWakeWord is the
|
||||
# free, local default (ONNX runtime); Porcupine is the premium engine.
|
||||
# openWakeWord's ONNX embedding model returns near-zero scores on macOS
|
||||
# ARM64 (dscripka/openWakeWord#336), so the wake word runs on the tflite
|
||||
# backend there. Upstream declares tflite-runtime for Linux only;
|
||||
# ai-edge-litert is the macOS equivalent, bridged in tools/wake_word.py.
|
||||
# It lives in its own feature because lazy-dep specs cannot carry PEP 508
|
||||
# environment markers (_spec_is_safe rejects ";"), so the platform gate is
|
||||
# applied by the caller instead.
|
||||
"wake.openwakeword.tflite": (
|
||||
"ai-edge-litert==2.1.6",
|
||||
),
|
||||
"wake.openwakeword": (
|
||||
"openwakeword==0.6.0",
|
||||
"onnxruntime==1.27.0",
|
||||
|
|
|
|||
|
|
@ -33,6 +33,7 @@ from __future__ import annotations
|
|||
|
||||
import logging
|
||||
import os
|
||||
import sys
|
||||
import threading
|
||||
import time
|
||||
from pathlib import Path
|
||||
|
|
@ -78,6 +79,56 @@ def _bundled_wakeword_path(framework: str = "onnx") -> str:
|
|||
return os.path.join(os.path.dirname(__file__), "wakewords", f"{_BUNDLED_MODEL_NAME}.{ext}")
|
||||
|
||||
|
||||
def _is_macos_arm64() -> bool:
|
||||
import platform
|
||||
|
||||
return sys.platform == "darwin" and platform.machine() == "arm64"
|
||||
|
||||
|
||||
def default_inference_framework() -> str:
|
||||
"""The openWakeWord backend to use on this platform.
|
||||
|
||||
openWakeWord's ONNX backend produces near-zero scores on macOS ARM64 — its
|
||||
shared *embedding* model is the broken stage (the melspectrogram front-end
|
||||
and the wake classifier both match tflite exactly). The detector arms, the
|
||||
microphone works, and no phrase can ever cross the threshold. Prefer the
|
||||
tflite backend there; ONNX stays the default everywhere else.
|
||||
|
||||
Upstream: https://github.com/dscripka/openWakeWord/issues/336
|
||||
"""
|
||||
return "tflite" if _is_macos_arm64() else "onnx"
|
||||
|
||||
|
||||
def ensure_tflite_runtime() -> bool:
|
||||
"""Make ``import tflite_runtime.interpreter`` resolve, returning success.
|
||||
|
||||
openWakeWord hardcodes that import but only declares ``tflite-runtime`` for
|
||||
``platform_system == "Linux"``; on macOS the equivalent wheel is
|
||||
``ai-edge-litert``. Alias the module so the upstream import succeeds. The
|
||||
alias is process-local — nothing is written to site-packages.
|
||||
"""
|
||||
try:
|
||||
import tflite_runtime.interpreter # noqa: F401
|
||||
|
||||
return True
|
||||
except ImportError:
|
||||
pass
|
||||
|
||||
try:
|
||||
from ai_edge_litert import interpreter as _litert # type: ignore[import-not-found]
|
||||
except ImportError:
|
||||
return False
|
||||
|
||||
import types
|
||||
|
||||
pkg = types.ModuleType("tflite_runtime")
|
||||
pkg.__path__ = [] # type: ignore[attr-defined] # mark as package
|
||||
sys.modules.setdefault("tflite_runtime", pkg)
|
||||
sys.modules["tflite_runtime.interpreter"] = _litert
|
||||
logger.debug("wake word: bridged tflite_runtime -> ai_edge_litert")
|
||||
return True
|
||||
|
||||
|
||||
def load_wake_word_config() -> Dict[str, Any]:
|
||||
"""Return the ``wake_word`` config section, shape-guarded to a dict."""
|
||||
try:
|
||||
|
|
@ -242,9 +293,32 @@ class _OpenWakeWordEngine(_Engine):
|
|||
|
||||
sub = cfg.get("openwakeword") if isinstance(cfg.get("openwakeword"), dict) else {}
|
||||
model_ref = str(sub.get("model") or _BUNDLED_MODEL_NAME).strip()
|
||||
framework = str(sub.get("inference_framework") or "onnx").strip().lower()
|
||||
framework = str(sub.get("inference_framework") or "").strip().lower()
|
||||
if not framework:
|
||||
framework = default_inference_framework()
|
||||
self._threshold = _sensitivity(cfg)
|
||||
|
||||
# openWakeWord silently downgrades tflite -> onnx when no tflite runtime
|
||||
# imports (model.py). On macOS ARM64 that lands on the backend whose
|
||||
# embedding model is broken, so the listener would arm and never fire.
|
||||
# Install + bridge the runtime first, and refuse the downgrade rather
|
||||
# than ship a dead ear.
|
||||
if framework == "tflite" and not ensure_tflite_runtime():
|
||||
# Same lazy-install contract as every other backend; the platform
|
||||
# gate lives here because dep specs can't carry PEP 508 markers.
|
||||
try:
|
||||
lazy_deps.ensure("wake.openwakeword.tflite", prompt=False)
|
||||
except Exception as e:
|
||||
logger.debug("wake word: tflite runtime install failed: %s", e)
|
||||
if not ensure_tflite_runtime():
|
||||
if _is_macos_arm64():
|
||||
raise RuntimeError(
|
||||
"The wake word needs the tflite backend on this Mac, but its "
|
||||
"runtime is missing. Install it with: pip install ai-edge-litert"
|
||||
)
|
||||
logger.warning("wake word: no tflite runtime available — falling back to onnx")
|
||||
framework = "onnx"
|
||||
|
||||
# Default (or explicit "hey_hermes") → the bundled model; a built-in name
|
||||
# or custom path is used as-is.
|
||||
if model_ref.lower() in _BUNDLED_MODEL_ALIASES:
|
||||
|
|
@ -519,16 +593,27 @@ def check_wake_word_requirements(cfg: Optional[Dict[str, Any]] = None) -> Dict[s
|
|||
key_ok = True
|
||||
hint = ""
|
||||
|
||||
# The tflite backend needs a runtime openWakeWord doesn't declare off Linux.
|
||||
# Report it as a real remediation instead of arming a detector that can't fire.
|
||||
tflite_ok = True
|
||||
if provider not in ("porcupine", "sherpa", "sherpa-onnx", "kws", "open"):
|
||||
sub = cfg.get("openwakeword") if isinstance(cfg.get("openwakeword"), dict) else {}
|
||||
framework = str(sub.get("inference_framework") or "").strip().lower() or default_inference_framework()
|
||||
if framework == "tflite":
|
||||
tflite_ok = ensure_tflite_runtime() or lazy_deps.is_available("wake.openwakeword.tflite") or lazy_ok
|
||||
|
||||
if provider == "porcupine" and not (os.getenv("PORCUPINE_ACCESS_KEY") or "").strip():
|
||||
key_ok = False
|
||||
hint = "Set PORCUPINE_ACCESS_KEY (free key at https://console.picovoice.ai)."
|
||||
elif not deps_ok and not lazy_ok:
|
||||
hint = lazy_deps.feature_install_command(feature) or ""
|
||||
elif not tflite_ok:
|
||||
hint = "The wake word needs the tflite runtime on this Mac: pip install ai-edge-litert"
|
||||
elif deps_ok and not audio_ok:
|
||||
hint = "Microphone capture needs sounddevice + numpy and a working audio device."
|
||||
|
||||
return {
|
||||
"available": key_ok and ((deps_ok and audio_ok) or (not deps_ok and lazy_ok)),
|
||||
"available": key_ok and tflite_ok and ((deps_ok and audio_ok) or (not deps_ok and lazy_ok)),
|
||||
"provider": provider,
|
||||
"deps_available": deps_ok,
|
||||
"audio_available": audio_ok,
|
||||
|
|
|
|||
33
uv.lock
generated
33
uv.lock
generated
|
|
@ -19,6 +19,24 @@ wheels = [
|
|||
{ url = "https://files.pythonhosted.org/packages/8f/ed/c284543c08aa443a4ef2c8bd120be51da8433dd174c01749b5d87c333f22/agent_client_protocol-0.9.0-py3-none-any.whl", hash = "sha256:06911500b51d8cb69112544e2be01fc5e7db39ef88fecbc3848c5c6f194798ee", size = 56850, upload-time = "2026-03-26T01:20:59.252Z" },
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "ai-edge-litert"
|
||||
version = "2.1.6"
|
||||
source = { registry = "https://pypi.org/simple" }
|
||||
dependencies = [
|
||||
{ name = "backports-strenum" },
|
||||
{ name = "flatbuffers" },
|
||||
{ name = "numpy" },
|
||||
{ name = "protobuf" },
|
||||
{ name = "tqdm" },
|
||||
{ name = "typing-extensions" },
|
||||
]
|
||||
wheels = [
|
||||
{ url = "https://files.pythonhosted.org/packages/b6/3d/41a85023e1c6cc76d895f3ba6ac7c22b0785db7e08d0827ebeb8a403eefc/ai_edge_litert-2.1.6-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:edf598814004e594b40c888f52cae59e950dbeffd821e83ba45d28db0a0aa3f5", size = 10031164, upload-time = "2026-07-01T21:42:50.986Z" },
|
||||
{ url = "https://files.pythonhosted.org/packages/52/d5/164aaf69f60f72b7076900ef1cc6153bf50d82cd15202bdf1239c0dbfb1c/ai_edge_litert-2.1.6-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:5adf0c9afde6151dc7f2989d039c800f3060d98d40bb5dfc95e426ad4eb3680b", size = 10033170, upload-time = "2026-07-01T21:42:53.208Z" },
|
||||
{ url = "https://files.pythonhosted.org/packages/ad/92/578b31c4c05afa9081a664cd86afe33c304fecd70de1c2a4d3a3b9ca51c9/ai_edge_litert-2.1.6-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:2e8a3f92fa407690189533bea8b64d49bb1b8a9e96f707ba27e1a64a9c3cc8cf", size = 10032950, upload-time = "2026-07-01T21:42:55.102Z" },
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "aiofiles"
|
||||
version = "24.1.0"
|
||||
|
|
@ -472,6 +490,15 @@ wheels = [
|
|||
{ url = "https://files.pythonhosted.org/packages/df/73/b6e24bd22e6720ca8ee9a85a0c4a2971af8497d8f3193fa05390cbd46e09/backoff-2.2.1-py3-none-any.whl", hash = "sha256:63579f9a0628e06278f7e47b7d7d5b6ce20dc65c5e96a6f3ca99a6adca0396e8", size = 15148, upload-time = "2022-10-05T19:19:30.546Z" },
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "backports-strenum"
|
||||
version = "1.3.1"
|
||||
source = { registry = "https://pypi.org/simple" }
|
||||
sdist = { url = "https://files.pythonhosted.org/packages/35/c7/2ed54c32fed313591ffb21edbd48db71e68827d43a61938e5a0bc2b6ec91/backports_strenum-1.3.1.tar.gz", hash = "sha256:77c52407342898497714f0596e86188bb7084f89063226f4ba66863482f42414", size = 7257, upload-time = "2023-12-09T14:36:40.937Z" }
|
||||
wheels = [
|
||||
{ url = "https://files.pythonhosted.org/packages/d6/50/56cf20e2ee5127b603b81d5a69580a1a325083e2b921aa8f067da83927c0/backports_strenum-1.3.1-py3-none-any.whl", hash = "sha256:cdcfe36dc897e2615dc793b7d3097f54d359918fc448754a517e6f23044ccf83", size = 8304, upload-time = "2023-12-09T14:36:39.905Z" },
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "base58"
|
||||
version = "2.1.1"
|
||||
|
|
@ -1721,6 +1748,7 @@ voice = [
|
|||
{ name = "sounddevice" },
|
||||
]
|
||||
wake = [
|
||||
{ name = "ai-edge-litert", marker = "sys_platform == 'darwin'" },
|
||||
{ name = "numpy" },
|
||||
{ name = "onnxruntime" },
|
||||
{ name = "openwakeword" },
|
||||
|
|
@ -1745,6 +1773,7 @@ youtube = [
|
|||
[package.metadata]
|
||||
requires-dist = [
|
||||
{ name = "agent-client-protocol", marker = "extra == 'acp'", specifier = "==0.9.0" },
|
||||
{ name = "ai-edge-litert", marker = "sys_platform == 'darwin' and extra == 'wake'", specifier = "==2.1.6" },
|
||||
{ name = "aiohttp", marker = "extra == 'homeassistant'", specifier = "==3.14.1" },
|
||||
{ name = "aiohttp", marker = "extra == 'matrix'", specifier = "==3.14.1" },
|
||||
{ name = "aiohttp", marker = "extra == 'messaging'", specifier = "==3.14.1" },
|
||||
|
|
@ -3942,7 +3971,7 @@ resolution-markers = [
|
|||
"python_full_version < '3.12'",
|
||||
]
|
||||
dependencies = [
|
||||
{ name = "numpy", marker = "python_full_version < '3.12'" },
|
||||
{ name = "numpy" },
|
||||
]
|
||||
sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
|
||||
wheels = [
|
||||
|
|
@ -3997,7 +4026,7 @@ resolution-markers = [
|
|||
"python_full_version == '3.12.*'",
|
||||
]
|
||||
dependencies = [
|
||||
{ name = "numpy", marker = "python_full_version >= '3.12'" },
|
||||
{ name = "numpy" },
|
||||
]
|
||||
sdist = { url = "https://files.pythonhosted.org/packages/a7/25/c2700dfaf6442b4effaa91af24ebce5dc9d31bb4a69706313aae70d72cd0/scipy-1.18.0.tar.gz", hash = "sha256:67b2ad2ad54c72ca6d04975a9b2df8c3638c34ddd5b28738e94fc2b57929d378", size = 30774447, upload-time = "2026-06-19T15:01:43.456Z" }
|
||||
wheels = [
|
||||
|
|
|
|||
|
|
@ -79,7 +79,7 @@ wake_word:
|
|||
start_new_session: true # start a fresh session on wake vs. continue the current one
|
||||
openwakeword:
|
||||
model: hey_hermes # bundled default; OR a built-in name OR a path to a custom .onnx/.tflite
|
||||
inference_framework: onnx # "onnx" | "tflite"
|
||||
inference_framework: "" # "" (auto) | "onnx" | "tflite"
|
||||
porcupine:
|
||||
keyword: jarvis # built-in keyword OR path to a custom .ppn
|
||||
```
|
||||
|
|
@ -87,6 +87,14 @@ wake_word:
|
|||
`sensitivity`, `phrase`, and `start_new_session` apply to both engines. The
|
||||
`openwakeword` and `porcupine` blocks select the actual detection model.
|
||||
|
||||
`inference_framework` picks the openWakeWord backend. Leave it empty (the
|
||||
default) to let Hermes choose per platform: **tflite on Apple Silicon**, onnx
|
||||
everywhere else. openWakeWord's onnx backend returns near-zero scores on macOS
|
||||
ARM64 ([openWakeWord#336](https://github.com/dscripka/openWakeWord/issues/336)),
|
||||
so a listener pinned to `onnx` there will arm, show as listening, and never
|
||||
fire. The tflite backend needs `ai-edge-litert` on macOS, which Hermes installs
|
||||
on demand alongside the other wake-word deps.
|
||||
|
||||
### Surfaces (CLI, TUI, GUI)
|
||||
|
||||
The wake word works in all three Hermes surfaces, and `surface` picks which one
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue