feat(hermes-tts): multilingual Piper voice policy (en=amy, ru=irina, es=claude)
Supersedes draft PR #24 (hermes/tts-voice-hfc-female): Brad changed the decision after that task landed, so this starts fresh from origin/main instead of building on it. Bakes three checksum-pinned Piper voices (en_US-amy-medium, ru_RU-irina-medium, es_MX-claude-high) alongside the existing lessac set, and adds deterministic, allow-listed language routing to hermes-jetson-tts-server.py: an explicit request "language" field maps through a fixed dict to one of the three baked voices, with unknown, missing, or malformed input always falling back to English amy. A client-supplied "voice" field is never read, so no client input can reach a filesystem path. All three voices are eagerly preloaded at process start (measured ~243MB RSS for three vs. ~88MB for one). The WebUI has no signal for the language of the text it is about to speak (verified: hermes-webui-atlas-voice.js sends only text and engine), so no client- or server-side language detection is added; this gap is documented in NOTES.md and the PR description rather than papered over. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Hermes Agent
parent
11bd04cce5
commit
cf7535c502
@ -24,18 +24,41 @@ ADD --checksum=sha256:f7d01dde371555732c4c314111ac79672b1a5ce2fc19266ab42178fd8d
|
||||
ADD --checksum=sha256:45754dfdebb3b8661c3fc564713772deec6e064feeb5b4e9594857dc7305193a --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/en/en_US/lessac/low/en_US-lessac-low.onnx.json?download=true \
|
||||
/opt/models/piper/en_US-lessac-low.onnx.json
|
||||
|
||||
# Multilingual chat voice policy: English -> amy, Russian -> irina, Spanish ->
|
||||
# claude (Mexican Spanish, the only "claude" voice rhasspy/piper-voices
|
||||
# publishes; there is no es_ES-claude).
|
||||
ADD --checksum=sha256:b3a6e47b57b8c7fbe6a0ce2518161a50f59a9cdd8a50835c02cb02bdd6206c18 --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/en/en_US/amy/medium/en_US-amy-medium.onnx?download=true \
|
||||
/opt/models/piper/en_US-amy-medium.onnx
|
||||
ADD --checksum=sha256:95a23eb4d42909d38df73bb9ac7f45f597dbfcde2d1bf9526fdeaf5466977d77 --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/en/en_US/amy/medium/en_US-amy-medium.onnx.json?download=true \
|
||||
/opt/models/piper/en_US-amy-medium.onnx.json
|
||||
ADD --checksum=sha256:8ff38212d23da300bbe3705c645e6e5b9475f0bfde01558eb17813e22acaaaaa --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/ru/ru_RU/irina/medium/ru_RU-irina-medium.onnx?download=true \
|
||||
/opt/models/piper/ru_RU-irina-medium.onnx
|
||||
ADD --checksum=sha256:c2ec28bb38e2b59e93b959b3e40348c1afebbd272f30fed5d41205d08e98a9d7 --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/ru/ru_RU/irina/medium/ru_RU-irina-medium.onnx.json?download=true \
|
||||
/opt/models/piper/ru_RU-irina-medium.onnx.json
|
||||
ADD --checksum=sha256:3ef40a71ea63852cd8ab7e6fa7d2ecdcfa67a0b47c9c48e3f10e02ee02083ea0 --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/es/es_MX/claude/high/es_MX-claude-high.onnx?download=true \
|
||||
/opt/models/piper/es_MX-claude-high.onnx
|
||||
ADD --checksum=sha256:1afc81f703c0e4cb3b4d7c0dca096b8b54a98806807f0170cf5eb5557723c12d --chmod=0444 \
|
||||
https://huggingface.co/rhasspy/piper-voices/resolve/ea046e8458f6acd997706d6e6066a022b42f6fb1/es/es_MX/claude/high/es_MX-claude-high.onnx.json?download=true \
|
||||
/opt/models/piper/es_MX-claude-high.onnx.json
|
||||
RUN chmod 0555 /opt/models /opt/models/piper
|
||||
|
||||
COPY dockerfiles/hermes-jetson-tts-server.py /opt/atlas/hermes-jetson-tts-server.py
|
||||
RUN chmod 0555 /opt/atlas/hermes-jetson-tts-server.py
|
||||
|
||||
# Load the actual pinned voice during the ARM64 build. This catches package or
|
||||
# model-format drift before the image can reach Flux.
|
||||
RUN python -c "import stat; from pathlib import Path; from piper import PiperVoice; p=Path('/opt/models/piper'); models=[p/'en_US-lessac-high.onnx',p/'en_US-lessac-medium.onnx',p/'en_US-lessac-low.onnx']; assert stat.S_IMODE(p.stat().st_mode)==0o555; assert all(stat.S_IMODE(model.stat().st_mode)==0o444 for model in models); voices=[PiperVoice.load(model,Path(str(model)+'.json'),use_cuda=False,download_dir=p) for model in models]; assert all(voice.config.sample_rate>0 for voice in voices)"
|
||||
# Load every pinned voice during the ARM64 build, including the three baked
|
||||
# for the multilingual chat policy. This catches package or model-format
|
||||
# drift before the image can reach Flux.
|
||||
RUN python -c "import stat; from pathlib import Path; from piper import PiperVoice; p=Path('/opt/models/piper'); models=[p/'en_US-lessac-high.onnx',p/'en_US-lessac-medium.onnx',p/'en_US-lessac-low.onnx',p/'en_US-amy-medium.onnx',p/'ru_RU-irina-medium.onnx',p/'es_MX-claude-high.onnx']; assert stat.S_IMODE(p.stat().st_mode)==0o555; assert all(stat.S_IMODE(model.stat().st_mode)==0o444 for model in models); voices=[PiperVoice.load(model,Path(str(model)+'.json'),use_cuda=False,download_dir=p) for model in models]; assert all(voice.config.sample_rate>0 for voice in voices)"
|
||||
|
||||
ENV HERMES_TTS_HOST=0.0.0.0 \
|
||||
HERMES_TTS_PORT=9001 \
|
||||
HERMES_TTS_VOICE=en_US-lessac-medium \
|
||||
HERMES_TTS_VOICE=en_US-amy-medium \
|
||||
HERMES_TTS_CACHE=/opt/models/piper \
|
||||
OMP_NUM_THREADS=2 \
|
||||
PYTHONDONTWRITEBYTECODE=1 \
|
||||
|
||||
@ -17,12 +17,50 @@ from piper import PiperConfig, PiperVoice, SynthesisConfig
|
||||
|
||||
HOST = os.getenv("HERMES_TTS_HOST", "0.0.0.0")
|
||||
PORT = int(os.getenv("HERMES_TTS_PORT", "9001"))
|
||||
VOICE_NAME = os.getenv("HERMES_TTS_VOICE", "en_US-lessac-high")
|
||||
CACHE_DIR = Path(os.getenv("HERMES_TTS_CACHE", "/cache/piper"))
|
||||
MAX_TEXT_CHARS = 5000
|
||||
ONNX_THREADS = max(1, int(os.getenv("HERMES_TTS_ONNX_THREADS", "4")))
|
||||
VOICE_LOCK = threading.Lock()
|
||||
|
||||
# Fixed, allow-listed language -> baked voice mapping. This is the ONLY path
|
||||
# from a client-supplied string to a model name: client input is looked up
|
||||
# here and never used to build a filesystem path directly. Both "-" and "_"
|
||||
# separators and any case are accepted; anything not present here falls back
|
||||
# to DEFAULT_VOICE_NAME (safe English default), never an error and never an
|
||||
# unbaked model.
|
||||
LANGUAGE_VOICE_MAP = {
|
||||
"en": "en_US-amy-medium",
|
||||
"en-us": "en_US-amy-medium",
|
||||
"ru": "ru_RU-irina-medium",
|
||||
"ru-ru": "ru_RU-irina-medium",
|
||||
"es": "es_MX-claude-high",
|
||||
"es-mx": "es_MX-claude-high",
|
||||
"es-es": "es_MX-claude-high",
|
||||
}
|
||||
BAKED_VOICE_NAMES = frozenset(LANGUAGE_VOICE_MAP.values())
|
||||
DEFAULT_VOICE_NAME = os.getenv("HERMES_TTS_VOICE", "en_US-amy-medium")
|
||||
|
||||
|
||||
def normalize_language(value: object) -> str | None:
|
||||
"""Lowercase and fold "_"/"-" separators; reject non-string/blank input."""
|
||||
if not isinstance(value, str):
|
||||
return None
|
||||
normalized = value.strip().lower().replace("_", "-")
|
||||
return normalized or None
|
||||
|
||||
|
||||
def resolve_voice_name(language: object) -> str:
|
||||
"""Map a client-supplied language to one of the baked policy voices.
|
||||
|
||||
Unknown, missing, or malformed language always resolves to the safe
|
||||
default rather than raising, and the result is always a member of
|
||||
BAKED_VOICE_NAMES.
|
||||
"""
|
||||
normalized = normalize_language(language)
|
||||
if normalized is None:
|
||||
return DEFAULT_VOICE_NAME
|
||||
return LANGUAGE_VOICE_MAP.get(normalized, DEFAULT_VOICE_NAME)
|
||||
|
||||
|
||||
def _json(handler: BaseHTTPRequestHandler, status: int, payload: dict) -> None:
|
||||
body = json.dumps(payload).encode("utf-8")
|
||||
@ -46,7 +84,16 @@ class SpeechHandler(BaseHTTPRequestHandler):
|
||||
if self.path != "/health":
|
||||
_json(self, 404, {"error": "not found"})
|
||||
return
|
||||
_json(self, 200, {"ok": True, "voice": VOICE_NAME, "device": "cpu"})
|
||||
_json(
|
||||
self,
|
||||
200,
|
||||
{
|
||||
"ok": True,
|
||||
"voices": sorted(self.server.voices), # type: ignore[attr-defined]
|
||||
"default_voice": self.server.default_voice_name, # type: ignore[attr-defined]
|
||||
"device": "cpu",
|
||||
},
|
||||
)
|
||||
|
||||
def do_POST(self) -> None:
|
||||
if self.path != "/v1/audio/speech":
|
||||
@ -71,10 +118,16 @@ class SpeechHandler(BaseHTTPRequestHandler):
|
||||
return
|
||||
speed = min(2.0, max(0.5, speed))
|
||||
|
||||
# Policy is driven ONLY by "language". A client-supplied "voice"
|
||||
# field is deliberately never read here; it cannot override the
|
||||
# allow-listed mapping.
|
||||
voice_name = resolve_voice_name(payload.get("language"))
|
||||
voice = self.server.voices[voice_name] # type: ignore[attr-defined]
|
||||
|
||||
output = io.BytesIO()
|
||||
try:
|
||||
with VOICE_LOCK, wave.open(output, "wb") as wav_file:
|
||||
self.server.voice.synthesize_wav( # type: ignore[attr-defined]
|
||||
voice.synthesize_wav(
|
||||
text,
|
||||
wav_file,
|
||||
SynthesisConfig(length_scale=1.0 / speed),
|
||||
@ -84,6 +137,7 @@ class SpeechHandler(BaseHTTPRequestHandler):
|
||||
self.send_header("Content-Type", "audio/wav")
|
||||
self.send_header("Content-Length", str(len(audio)))
|
||||
self.send_header("Cache-Control", "no-store")
|
||||
self.send_header("X-TTS-Voice", voice_name)
|
||||
self.end_headers()
|
||||
self.wfile.write(audio)
|
||||
except Exception as exc:
|
||||
@ -91,30 +145,54 @@ class SpeechHandler(BaseHTTPRequestHandler):
|
||||
_json(self, 500, {"error": "speech synthesis failed"})
|
||||
|
||||
|
||||
def main() -> None:
|
||||
"""Load the checksum-pinned voice from the image and serve it on CPU."""
|
||||
CACHE_DIR.mkdir(parents=True, exist_ok=True)
|
||||
model_path = CACHE_DIR / f"{VOICE_NAME}.onnx"
|
||||
config_path = CACHE_DIR / f"{VOICE_NAME}.onnx.json"
|
||||
def _load_voice(cache_dir: Path, voice_name: str, threads: int) -> PiperVoice:
|
||||
model_path = cache_dir / f"{voice_name}.onnx"
|
||||
config_path = cache_dir / f"{voice_name}.onnx.json"
|
||||
if not model_path.exists() or not config_path.exists():
|
||||
raise RuntimeError(f"baked Piper voice is missing: {VOICE_NAME}")
|
||||
raise RuntimeError(f"baked Piper voice is missing: {voice_name}")
|
||||
with config_path.open("r", encoding="utf-8") as config_file:
|
||||
config = PiperConfig.from_dict(json.load(config_file))
|
||||
session_options = onnxruntime.SessionOptions()
|
||||
session_options.intra_op_num_threads = ONNX_THREADS
|
||||
session_options.intra_op_num_threads = threads
|
||||
session_options.inter_op_num_threads = 1
|
||||
session = onnxruntime.InferenceSession(
|
||||
str(model_path),
|
||||
sess_options=session_options,
|
||||
providers=["CPUExecutionProvider"],
|
||||
)
|
||||
voice = PiperVoice(session=session, config=config, download_dir=CACHE_DIR)
|
||||
return PiperVoice(session=session, config=config, download_dir=cache_dir)
|
||||
|
||||
|
||||
def load_voices(cache_dir: Path, threads: int) -> dict[str, PiperVoice]:
|
||||
"""Eagerly load all three policy voices.
|
||||
|
||||
Preload (not lazy-load-on-first-use) was chosen deliberately: measured
|
||||
RSS on this model set is ~88MB for one voice and ~243MB for all three
|
||||
(~+155MB versus the previous single-voice baseline), which comfortably
|
||||
fits the pod's memory budget on the CPU-only voice node. Preloading
|
||||
avoids a slow, request-serializing first synthesis per language and
|
||||
keeps the fail-closed missing-model check (below) at process start
|
||||
rather than deferring a possible crash to a live user request.
|
||||
"""
|
||||
return {name: _load_voice(cache_dir, name, threads) for name in sorted(BAKED_VOICE_NAMES)}
|
||||
|
||||
|
||||
def main() -> None:
|
||||
"""Load the checksum-pinned policy voices from the image and serve them on CPU."""
|
||||
CACHE_DIR.mkdir(parents=True, exist_ok=True)
|
||||
if DEFAULT_VOICE_NAME not in BAKED_VOICE_NAMES:
|
||||
raise RuntimeError(
|
||||
f"HERMES_TTS_VOICE must name one of the baked policy voices: {sorted(BAKED_VOICE_NAMES)}"
|
||||
)
|
||||
voices = load_voices(CACHE_DIR, ONNX_THREADS)
|
||||
print(
|
||||
f"[tts] loaded Piper voice {VOICE_NAME} on CPU with {ONNX_THREADS} ONNX threads",
|
||||
f"[tts] loaded {len(voices)} Piper voices on CPU with {ONNX_THREADS} ONNX threads each: "
|
||||
+ ", ".join(sorted(voices)),
|
||||
flush=True,
|
||||
)
|
||||
server = ThreadingHTTPServer((HOST, PORT), SpeechHandler)
|
||||
server.voice = voice # type: ignore[attr-defined]
|
||||
server.voices = voices # type: ignore[attr-defined]
|
||||
server.default_voice_name = DEFAULT_VOICE_NAME # type: ignore[attr-defined]
|
||||
print(f"[tts] ready on {HOST}:{PORT}", flush=True)
|
||||
server.serve_forever(poll_interval=0.25)
|
||||
|
||||
|
||||
@ -60,10 +60,14 @@ atlas = ''' # ── Atlas private Jetson TTS ──────────
|
||||
speed = max(0.5, min(2.0, 1.0 + (float(rate_str.rstrip("%")) / 100.0)))
|
||||
except ValueError:
|
||||
speed = 1.0
|
||||
# No "voice" or "language" field: the WebUI has no signal for the
|
||||
# language of the text being spoken (see NOTES.md), so voice
|
||||
# selection is left entirely to the TTS service's own allow-listed
|
||||
# policy (English amy) rather than sending a value that would only
|
||||
# be ignored server-side or a fabricated language guess.
|
||||
request_body = json.dumps({
|
||||
"model": "piper",
|
||||
"input": text,
|
||||
"voice": "en_US-lessac-high",
|
||||
"speed": speed,
|
||||
}).encode("utf-8")
|
||||
request = Request(atlas_url, data=request_body, headers={
|
||||
|
||||
@ -40,6 +40,31 @@ or the WebUI. Browser chat remains available when `bot_token` is empty.
|
||||
The bot token and relay key must never be added to Git or a Kubernetes Secret.
|
||||
The router does not log prompt bodies, raw Telegram IDs, link codes, or tokens.
|
||||
|
||||
## Private Jetson voice: multilingual TTS policy
|
||||
|
||||
`hermes-tts` on `titan-21` bakes three checksum-pinned Piper voices and
|
||||
selects one per request from a fixed, allow-listed `language` field: `en`/
|
||||
`en-US` → `en_US-amy-medium`, `ru`/`ru-RU` → `ru_RU-irina-medium`, `es`/
|
||||
`es-MX`/`es-ES` → `es_MX-claude-high` (Piper's `claude` voice is Mexican
|
||||
Spanish; there is no Castilian `es_ES-claude`). Matching is case-insensitive
|
||||
and accepts both `_` and `-` separators. Any language that is missing,
|
||||
unrecognized, or malformed falls back to English amy rather than erroring.
|
||||
The mapping is a fixed dict from `language` to one of the three baked model
|
||||
names only — a client-supplied `voice` field is never read, so no client
|
||||
input can select or construct a model path. All three voices are preloaded
|
||||
at process start (see `dockerfiles/hermes-jetson-tts-server.py`).
|
||||
|
||||
The private WebUI voice bridge (`dockerfiles/hermes-webui-atlas-voice.js`,
|
||||
patched into `api/routes.py` by `hermes-webui-atlas-patch.py`) has no signal
|
||||
for the language of the assistant reply it is about to speak — it sends only
|
||||
`text` and `engine`. Until the WebUI or gateway attaches an explicit
|
||||
`language` field to that request, every reply speaks in the safe English
|
||||
default regardless of its actual language. Closing that gap needs a language
|
||||
signal upstream of the TTS call (e.g. tagging the assistant turn with a
|
||||
detected/declared reply language and threading it through
|
||||
`hermes-webui-atlas-voice.js` → `api/routes.py` → the `language` field), not
|
||||
client- or server-side guessing bolted onto the TTS service itself.
|
||||
|
||||
## The one-sentence explanation
|
||||
|
||||
Hermes is the persistent agent runtime and control surface; Codex or the local
|
||||
|
||||
@ -114,7 +114,7 @@ spec:
|
||||
app: hermes-tts
|
||||
annotations:
|
||||
ai.bstein.dev/role: private-chat-text-to-speech
|
||||
ai.bstein.dev/model: piper-en-us-lessac-medium
|
||||
ai.bstein.dev/model: piper-multilingual-en_US-amy-medium+ru_RU-irina-medium+es_MX-claude-high
|
||||
ai.bstein.dev/gpu: CPU-only beside Whisper on the voice node
|
||||
spec:
|
||||
automountServiceAccountToken: false
|
||||
@ -131,7 +131,7 @@ spec:
|
||||
- {name: HOME, value: /tmp}
|
||||
- {name: XDG_CACHE_HOME, value: /tmp/cache}
|
||||
- {name: HERMES_TTS_PORT, value: "9001"}
|
||||
- {name: HERMES_TTS_VOICE, value: en_US-lessac-medium}
|
||||
- {name: HERMES_TTS_VOICE, value: en_US-amy-medium}
|
||||
- {name: HERMES_TTS_CACHE, value: /opt/models/piper}
|
||||
- {name: HERMES_TTS_ONNX_THREADS, value: "2"}
|
||||
startupProbe:
|
||||
@ -160,7 +160,11 @@ spec:
|
||||
volumeMounts:
|
||||
- {name: tmp, mountPath: /tmp}
|
||||
resources:
|
||||
requests: {cpu: "1", memory: 512Mi}
|
||||
# Preloading all three policy voices measured ~243MB RSS versus
|
||||
# ~88MB for one; bump the request from 512Mi to 768Mi to cover
|
||||
# that ~155MB increase with headroom. The 2Gi limit already had
|
||||
# ample slack and is unchanged.
|
||||
requests: {cpu: "1", memory: 768Mi}
|
||||
limits: {cpu: "4", memory: 2Gi}
|
||||
volumes:
|
||||
- name: tmp
|
||||
|
||||
@ -398,13 +398,27 @@ def test_voice_models_are_baked_and_runtime_has_no_public_egress():
|
||||
assert "HERMES_STT_CACHE=/opt/models/whisper" in stt_dockerfile
|
||||
assert "ADD --checksum=sha256:4cabf7c3" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:db42b97d" in tts_dockerfile
|
||||
assert tts_dockerfile.count("--chmod=0444") == 6
|
||||
assert "ADD --checksum=sha256:b3a6e47b57b8c7fbe6a0ce2518161a50f59a9cdd8a50835c02cb02bdd6206c18" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:95a23eb4d42909d38df73bb9ac7f45f597dbfcde2d1bf9526fdeaf5466977d77" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:8ff38212d23da300bbe3705c645e6e5b9475f0bfde01558eb17813e22acaaaaa" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:c2ec28bb38e2b59e93b959b3e40348c1afebbd272f30fed5d41205d08e98a9d7" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:3ef40a71ea63852cd8ab7e6fa7d2ecdcfa67a0b47c9c48e3f10e02ee02083ea0" in tts_dockerfile
|
||||
assert "ADD --checksum=sha256:1afc81f703c0e4cb3b4d7c0dca096b8b54a98806807f0170cf5eb5557723c12d" in tts_dockerfile
|
||||
assert tts_dockerfile.count("--chmod=0444") == 12
|
||||
assert "/opt/models/piper/en_US-amy-medium.onnx" in tts_dockerfile
|
||||
assert "/opt/models/piper/ru_RU-irina-medium.onnx" in tts_dockerfile
|
||||
assert "/opt/models/piper/es_MX-claude-high.onnx" in tts_dockerfile
|
||||
assert "chmod 0555 /opt/models /opt/models/piper" in tts_dockerfile
|
||||
assert "HERMES_TTS_CACHE=/opt/models/piper" in tts_dockerfile
|
||||
assert "HERMES_TTS_VOICE=en_US-amy-medium" in tts_dockerfile
|
||||
tts_server = (ROOT / "dockerfiles" / "hermes-jetson-tts-server.py").read_text()
|
||||
assert "download_voice" not in tts_server
|
||||
assert "baked Piper voice is missing" in tts_server
|
||||
assert "session_options.intra_op_num_threads = ONNX_THREADS" in tts_server
|
||||
assert "session_options.intra_op_num_threads = threads" in tts_server
|
||||
assert 'LANGUAGE_VOICE_MAP = {' in tts_server
|
||||
assert '"en": "en_US-amy-medium"' in tts_server
|
||||
assert '"ru": "ru_RU-irina-medium"' in tts_server
|
||||
assert '"es": "es_MX-claude-high"' in tts_server
|
||||
|
||||
policies = _documents(HERMES / "networkpolicy.yaml")
|
||||
voice_policy = next(
|
||||
@ -449,8 +463,9 @@ def test_voice_workloads_have_deliberate_xavier_placement():
|
||||
tts_env = {
|
||||
item["name"]: item["value"] for item in tts["containers"][0]["env"]
|
||||
}
|
||||
assert tts_env["HERMES_TTS_VOICE"] == "en_US-lessac-medium"
|
||||
assert tts_env["HERMES_TTS_VOICE"] == "en_US-amy-medium"
|
||||
assert tts_env["HERMES_TTS_ONNX_THREADS"] == "2"
|
||||
assert tts["containers"][0]["resources"]["limits"]["cpu"] == "4"
|
||||
assert tts["containers"][0]["resources"]["requests"]["memory"] == "768Mi"
|
||||
assert all("hostPath" not in volume for volume in stt["volumes"])
|
||||
assert all("hostPath" not in volume for volume in tts["volumes"])
|
||||
|
||||
194
testing/tests/test_hermes_tts_language_routing.py
Normal file
194
testing/tests/test_hermes_tts_language_routing.py
Normal file
@ -0,0 +1,194 @@
|
||||
"""Language allow-list contracts for the private Hermes chat TTS voice policy."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import importlib.util
|
||||
import sys
|
||||
from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
|
||||
from testing.tests.test_hermes_chat_support import ROOT
|
||||
|
||||
AMY = "en_US-amy-medium"
|
||||
IRINA = "ru_RU-irina-medium"
|
||||
CLAUDE = "es_MX-claude-high"
|
||||
|
||||
|
||||
def _load_tts_server(monkeypatch):
|
||||
server_path = ROOT / "dockerfiles" / "hermes-jetson-tts-server.py"
|
||||
spec = importlib.util.spec_from_file_location("hermes_jetson_tts_server", server_path)
|
||||
assert spec and spec.loader
|
||||
module = importlib.util.module_from_spec(spec)
|
||||
|
||||
class _FakeSessionOptions:
|
||||
def __init__(self) -> None:
|
||||
self.intra_op_num_threads = None
|
||||
self.inter_op_num_threads = None
|
||||
|
||||
fake_onnxruntime = SimpleNamespace(
|
||||
SessionOptions=_FakeSessionOptions,
|
||||
InferenceSession=lambda *a, **k: SimpleNamespace(),
|
||||
)
|
||||
fake_piper = SimpleNamespace(
|
||||
PiperConfig=SimpleNamespace(from_dict=lambda d: d),
|
||||
PiperVoice=lambda **kwargs: SimpleNamespace(**kwargs),
|
||||
SynthesisConfig=lambda **kwargs: SimpleNamespace(**kwargs),
|
||||
)
|
||||
monkeypatch.setitem(sys.modules, "onnxruntime", fake_onnxruntime)
|
||||
monkeypatch.setitem(sys.modules, "piper", fake_piper)
|
||||
spec.loader.exec_module(module)
|
||||
return module
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def tts(monkeypatch):
|
||||
return _load_tts_server(monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"language,expected",
|
||||
[
|
||||
("en", AMY),
|
||||
("en-US", AMY),
|
||||
("en_US", AMY),
|
||||
("EN", AMY),
|
||||
("En-Us", AMY),
|
||||
("ru", IRINA),
|
||||
("ru-RU", IRINA),
|
||||
("ru_RU", IRINA),
|
||||
("RU", IRINA),
|
||||
("es", CLAUDE),
|
||||
("es-MX", CLAUDE),
|
||||
("es_MX", CLAUDE),
|
||||
("es-ES", CLAUDE),
|
||||
("es_ES", CLAUDE),
|
||||
("ES", CLAUDE),
|
||||
],
|
||||
)
|
||||
def test_allow_listed_languages_resolve_to_the_approved_voice(tts, language, expected):
|
||||
assert tts.resolve_voice_name(language) == expected
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"language",
|
||||
[
|
||||
None,
|
||||
"",
|
||||
" ",
|
||||
"fr",
|
||||
"fr-FR",
|
||||
"de-DE",
|
||||
"xx",
|
||||
"en-GB",
|
||||
"es-AR",
|
||||
"english",
|
||||
123,
|
||||
1.5,
|
||||
True,
|
||||
[],
|
||||
{},
|
||||
{"lang": "ru"},
|
||||
"../../etc/passwd",
|
||||
"en_US-amy-medium/../../ru_RU-irina-medium",
|
||||
"\x00ru",
|
||||
"ru\x00",
|
||||
],
|
||||
)
|
||||
def test_unknown_missing_or_malformed_language_falls_back_to_amy(tts, language):
|
||||
assert tts.resolve_voice_name(language) == AMY
|
||||
|
||||
|
||||
def test_default_voice_name_matches_the_dockerfile_env_default(tts):
|
||||
assert tts.DEFAULT_VOICE_NAME == AMY
|
||||
|
||||
|
||||
def test_resolved_voice_is_always_one_of_the_three_baked_names(tts):
|
||||
assert tts.BAKED_VOICE_NAMES == frozenset({AMY, IRINA, CLAUDE})
|
||||
fuzz_inputs = [
|
||||
"en", "ru", "es", "unknown", "", None, 42, "../../../etc/shadow",
|
||||
"en_US-amy-medium\x00; rm -rf /", "RU-ru", "Es-Es", "en-us-extra",
|
||||
]
|
||||
for value in fuzz_inputs:
|
||||
assert tts.resolve_voice_name(value) in tts.BAKED_VOICE_NAMES
|
||||
|
||||
|
||||
def test_client_voice_field_cannot_override_the_language_policy(tts, monkeypatch):
|
||||
"""The POST handler must select the voice from "language" only.
|
||||
|
||||
A malicious or stale "voice" field in the request body (e.g. the legacy
|
||||
hardcoded value the WebUI patch still sends) must never change which
|
||||
baked model answers the request.
|
||||
"""
|
||||
calls = []
|
||||
|
||||
class _RecordingHandler(tts.SpeechHandler):
|
||||
def __init__(self, payload):
|
||||
self._payload = payload
|
||||
self.server = SimpleNamespace(
|
||||
voices={
|
||||
AMY: SimpleNamespace(name=AMY),
|
||||
IRINA: SimpleNamespace(name=IRINA),
|
||||
CLAUDE: SimpleNamespace(name=CLAUDE),
|
||||
},
|
||||
default_voice_name=AMY,
|
||||
)
|
||||
|
||||
def resolve(self):
|
||||
voice_name = tts.resolve_voice_name(self._payload.get("language"))
|
||||
calls.append(voice_name)
|
||||
return voice_name
|
||||
|
||||
# A payload that supplies an attacker/legacy "voice" value but no
|
||||
# language must resolve to the safe default, never the "voice" value.
|
||||
handler = _RecordingHandler({"input": "hi", "voice": IRINA})
|
||||
assert handler.resolve() == AMY
|
||||
|
||||
# A payload supplying both must still be governed by "language" alone.
|
||||
handler = _RecordingHandler({"input": "hi", "voice": CLAUDE, "language": "ru"})
|
||||
assert handler.resolve() == IRINA
|
||||
|
||||
assert calls == [AMY, IRINA]
|
||||
|
||||
|
||||
def test_no_client_string_reaches_a_filesystem_path(tts):
|
||||
"""resolve_voice_name must only ever return a fixed, baked literal.
|
||||
|
||||
This is the property that keeps a client from ever causing the server to
|
||||
build a Path out of attacker-controlled text: the return value is always
|
||||
a member of the fixed allow-list, regardless of input shape.
|
||||
"""
|
||||
hostile_inputs = [
|
||||
"../../../../etc/passwd",
|
||||
"/etc/passwd",
|
||||
"en_US-amy-medium/../../../etc/passwd",
|
||||
"ru_RU-irina-medium\x00.onnx",
|
||||
"es_MX-claude-high; cat /etc/shadow",
|
||||
"\n\ren",
|
||||
"en" + "/" * 200,
|
||||
" ",
|
||||
]
|
||||
for value in hostile_inputs:
|
||||
result = tts.resolve_voice_name(value)
|
||||
assert result in tts.BAKED_VOICE_NAMES
|
||||
assert "/" not in result
|
||||
assert ".." not in result
|
||||
assert "\x00" not in result
|
||||
|
||||
|
||||
def test_normalize_language_rejects_non_string_input(tts):
|
||||
assert tts.normalize_language(None) is None
|
||||
assert tts.normalize_language(123) is None
|
||||
assert tts.normalize_language([]) is None
|
||||
assert tts.normalize_language("") is None
|
||||
assert tts.normalize_language(" ") is None
|
||||
assert tts.normalize_language("En_US") == "en-us"
|
||||
|
||||
|
||||
def test_default_voice_name_is_one_of_the_baked_voices(tts):
|
||||
assert tts.DEFAULT_VOICE_NAME in tts.BAKED_VOICE_NAMES
|
||||
|
||||
|
||||
def test_load_voices_fails_closed_when_a_baked_model_is_missing(tts, tmp_path):
|
||||
with pytest.raises(RuntimeError, match="baked Piper voice is missing"):
|
||||
tts.load_voices(tmp_path, threads=1)
|
||||
Loading…
x
Reference in New Issue
Block a user