kokoro-tts

modules/containers/kokoro-tts.nix

@@ -0,0 +1,116 @@
+{
+  lib,
+  pkgs,
+  ...
+}:
+
+let
+  # The image is built here rather than pulled: zaakirio/kokoro-ru is a
+  # Hugging Face repo, not a published OCI image, and its Russian G2P has to be
+  # driven through the repo's own ru_g2p.py.
+  #
+  # The build context goes through the store so the image is pinned to the
+  # config revision: edit a file, `nixos-rebuild`, and the unit below rebuilds
+  # and restarts. Reading the context off a checkout at runtime would leave the
+  # running container untraceable back to any config.
+  source = pkgs.linkFarm "kokoro-tts-source" [
+    {
+      name = "Dockerfile";
+      path = toString ./kokoro-tts/Dockerfile;
+    }
+    {
+      name = "app.py";
+      path = toString ./kokoro-tts/app.py;
+    }
+    {
+      name = "fetch_assets.py";
+      path = toString ./kokoro-tts/fetch_assets.py;
+    }
+    {
+      name = "requirements.txt";
+      path = toString ./kokoro-tts/requirements.txt;
+    }
+  ];
+
+  image = "localhost/kokoro-tts:latest";
+
+  # Unchanged from the silero module, so whatever already points at
+  # http://127.0.0.1:9898/v1 keeps working without edits.
+  hostPort = 9898;
+  containerPort = 8000;
+in
+{
+  config = {
+    virtualisation = {
+      podman = {
+        enable = true;
+
+        autoPrune = {
+          enable = true;
+          flags = [ "--all" ];
+        };
+
+        dockerCompat = true;
+      };
+
+      oci-containers = {
+        backend = "podman";
+
+        containers.kokoro-tts = {
+          image = image;
+
+          ports = [
+            "127.0.0.1:${toString hostPort}:${toString containerPort}"
+          ];
+
+          environment = {
+            # Inference is CPU-bound and already threaded inside torch; these
+            # keep it from oversubscribing a small machine.
+            KOKORO_THREADS = "4";
+            OMP_NUM_THREADS = "4";
+            MKL_NUM_THREADS = "4";
+            TZ = "Europe/Moscow";
+          };
+
+          # No volumes: the checkpoints, the acute-aware espeak data and
+          # ruaccent's ONNX models are all baked into the image, so the
+          # container needs neither a host directory nor the network to start.
+          log-driver = "journald";
+        };
+      };
+    };
+
+    systemd = {
+      services = {
+        # Runs before the container. BuildKit caches the expensive layers, so
+        # on every boot after the first this is a no-op that still verifies the
+        # image exists.
+        "podman-build-kokoro-tts" = {
+          path = [ pkgs.podman ];
+
+          serviceConfig = {
+            Type = "oneshot";
+            RemainAfterExit = true;
+            # First build pulls torch wheels plus ~700 MB of weights.
+            TimeoutSec = 3600;
+          };
+
+          script = ''
+            podman build -t ${image} ${source}
+          '';
+
+          wantedBy = [ "multi-user.target" ];
+        };
+
+        "podman-kokoro-tts" = {
+          # The image does not exist until the build above ran, and a `latest`
+          # tag must be re-pulled on rebuild, so ordering has to be explicit.
+          after = [ "podman-build-kokoro-tts.service" ];
+          requires = [ "podman-build-kokoro-tts.service" ];
+          serviceConfig.Restart = lib.mkOverride 90 "always";
+          wantedBy = [ "multi-user.target" ];
+        };
+      };
+    };
+  };
+}

modules/containers/kokoro-tts/Dockerfile

@@ -0,0 +1,56 @@
+FROM python:3.12-slim-bookworm
+
+# Pinned, not "main": a rebuild that only touched the Nix module must not
+# silently pick up different weights. Bump these deliberately.
+ARG KOKORO_RU_REPO=zaakirio/kokoro-ru
+ARG KOKORO_RU_REVISION=d649c57b239b18c4c384378127cbf01dba039bc1
+# Trim to "sveta" to halve the image: masha shares her checkpoint and dima is
+# a second 327 MB one.
+ARG KOKORO_RU_VOICES=sveta,masha,dima
+
+ENV PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    HF_HUB_DISABLE_TELEMETRY=1 \
+    HF_HUB_DISABLE_SYMLINKS_WARNING=1 \
+    KOKORO_RU_REPO=${KOKORO_RU_REPO} \
+    KOKORO_RU_REVISION=${KOKORO_RU_REVISION} \
+    KOKORO_RU_VOICES=${KOKORO_RU_VOICES} \
+    KOKORO_MODEL_DIR=/app/kokoro-ru \
+    KOKORO_THREADS=4 \
+    OMP_NUM_THREADS=4 \
+    MKL_NUM_THREADS=4 \
+    TZ=Europe/Moscow
+
+WORKDIR /app
+
+# libgomp1 is torch's OpenMP runtime. espeak-ng comes from the espeakng-loader
+# wheel rather than the distro package because the model needs its own
+# recompiled ru_dict, and libsndfile is absent because WAV/PCM are written with
+# stdlib `wave` while every other format goes through imageio-ffmpeg.
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends libgomp1 \
+    && rm -rf /var/lib/apt/lists/*
+
+# CPU-only torch from its own index: the default PyPI wheel drags in ~2.5 GB of
+# CUDA libraries for a machine that has no GPU.
+RUN pip install --index-url https://download.pytorch.org/whl/cpu torch
+
+COPY requirements.txt ./
+RUN pip install -r requirements.txt
+
+COPY app.py fetch_assets.py ./
+
+# Bakes the checkpoints, the acute-aware espeak data and ruaccent's ONNX models
+# into the layer, which is what lets the container start with no network and no
+# writable volume.
+RUN python fetch_assets.py
+
+EXPOSE 8000
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=180s --retries=3 \
+    CMD ["python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/healthz', timeout=4)"]
+
+# No workers: the model is a shared in-process singleton, so a second worker
+# would only mean a second copy of ~2 GB of weights.
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "8000", "--workers", "1"]

modules/containers/kokoro-tts/app.py

@@ -0,0 +1,430 @@
+"""OpenAI-compatible TTS API backed by zaakirio/kokoro-ru.
+
+The model itself is language-blind: phoneme ids in, 24 kHz audio out. All the
+Russian lives in the G2P front-end, and the one that matters is kokoro-ru's own
+`ru_g2p.py` — RUAccent resolves lexical stress, ё and homographs, then an
+acute-aware espeak-ng phonemizer turns that into IPA. Stock misaki Russian is
+espeak-only and gets stress wrong often enough that the model reads as
+non-native (measured 27% vs 22% round-trip WER, per the model card).
+
+So: text -> RuG2P.phonemize -> KModel(ipa, voicepack[len(ipa) - 1]) -> waveform.
+
+Endpoints
+    POST /v1/audio/speech   OpenAI text-to-speech
+    GET  /v1/models         OpenAI model list
+    GET  /v1/voices         voice inventory (extension, not part of OpenAI)
+    GET  /healthz           readiness, 503 until the model is loaded
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+import os
+import re
+import subprocess
+import sys
+import threading
+import wave
+from contextlib import asynccontextmanager
+from pathlib import Path
+from typing import TYPE_CHECKING, Literal
+
+import numpy as np
+from fastapi import FastAPI
+from fastapi.responses import JSONResponse, Response
+from pydantic import BaseModel, ConfigDict, Field
+
+if TYPE_CHECKING:  # torch is imported lazily so /healthz answers during boot
+    import torch
+
+MODEL_ID = "kokoro-ru"
+SAMPLE_RATE = 24000
+MODEL_DIR = Path(os.environ.get("KOKORO_MODEL_DIR", "/app/kokoro-ru"))
+DEFAULT_VOICE = os.environ.get("KOKORO_DEFAULT_VOICE", "sveta")
+THREADS = int(os.environ.get("KOKORO_THREADS", os.cpu_count() or 4))
+# 2026-07-29, when the kokoro-ru revision we pin was published. Clients that
+# cache on this treat any change as a new model, so it must stay stable.
+MODEL_CREATED = 1785353253
+
+# voice -> (checkpoint stem, gender). The checkpoint carries the timbre and the
+# voicepack the identity, which is why sveta and masha share one file.
+VOICE_SPECS: dict[str, tuple[str, str]] = {
+    "sveta": ("kokoro-ru-v2-base", "female"),
+    "masha": ("kokoro-ru-v2-base", "female"),
+    "dima": ("kokoro-ru-v2-dima", "male"),
+}
+
+# Clients that ship the OpenAI voice list (alloy, nova, echo, ...) send those
+# names unless the user overrides them, so map them onto the three we have.
+VOICE_ALIASES: dict[str, str] = {
+    "alloy": "sveta",
+    "ash": "sveta",
+    "ballad": "sveta",
+    "verse": "sveta",
+    "marin": "sveta",
+    "coral": "masha",
+    "sage": "masha",
+    "shimmer": "masha",
+    "cedar": "masha",
+    "echo": "dima",
+    "fable": "dima",
+    "onyx": "dima",
+}
+
+CONTENT_TYPES = {
+    "wav": "audio/wav",
+    "mp3": "audio/mpeg",
+    "opus": "audio/ogg",
+    "aac": "audio/aac",
+    "flac": "audio/flac",
+    "pcm": "audio/pcm",
+}
+
+# Everything except wav and pcm goes through ffmpeg; those two are byte-exact
+# from the stdlib and need no encoder at all.
+FFMPEG_ARGS = {
+    "mp3": ["-c:a", "libmp3lame", "-q:a", "2"],
+    "opus": ["-c:a", "libopus", "-b:a", "64k"],
+    "aac": ["-c:a", "aac", "-b:a", "128k"],
+    "flac": ["-c:a", "flac"],
+}
+FFMPEG_CONTAINERS = {"mp3": "mp3", "opus": "ogg", "aac": "adts", "flac": "flac"}
+
+# Kokoro's Albert context is 510 tokens and KModel.forward asserts
+# len(ids) + 2 <= 510, so 508 phonemes is the hard ceiling per forward pass.
+MAX_PHONEMES = 508
+# Roughly 300 characters of Russian lands near 400 phonemes, comfortably under
+# the ceiling, and keeps a chunk short enough that a bad sentence is a short
+# chunk.
+CHUNK_CHARS = 300
+# Silence inserted between chunks. Without it the concatenation clicks at every
+# boundary because each forward pass starts and ends on a zero crossing.
+CHUNK_GAP_S = 0.08
+
+_SENTENCE_SPLIT = re.compile(r"(?<=[.!?…])\s+")
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+log = logging.getLogger("kokoro-ru")
+
+
+def split_text(text: str, budget: int = CHUNK_CHARS) -> list[str]:
+    """Split into sentence-bounded chunks, hard-cutting only as a last resort.
+
+    Phonemizing per sentence rather than per paragraph keeps RUAccent's stress
+    decisions local and gives the model a reset point at every full stop.
+    """
+    chunks: list[str] = []
+    current = ""
+    for sentence in _SENTENCE_SPLIT.split(text.strip()):
+        sentence = sentence.strip()
+        while len(sentence) > budget:
+            if current:
+                chunks.append(current)
+                current = ""
+            chunks.append(sentence[:budget])
+            sentence = sentence[budget:].strip()
+        if not sentence:
+            continue
+        if len(current) + len(sentence) + 1 > budget:
+            # Guarded: when the first sentence fills the budget exactly, or the
+            # previous one was hard-cut down to nothing, `current` is empty and
+            # a bare append would queue a zero-length chunk.
+            if current:
+                chunks.append(current)
+            current = sentence
+        else:
+            current = f"{current} {sentence}".strip()
+    if current:
+        chunks.append(current)
+    return chunks
+
+
+def split_phonemes(ps: str, limit: int = MAX_PHONEMES) -> list[str]:
+    """Cut an over-long phoneme string on word boundaries."""
+    if len(ps) <= limit:
+        return [ps]
+    parts: list[str] = []
+    rest = ps
+    while len(rest) > limit:
+        cut = rest.rfind(" ", 0, limit)
+        if cut <= 0:
+            cut = limit
+        parts.append(rest[:cut].strip())
+        rest = rest[cut:].strip()
+    if rest:
+        parts.append(rest)
+    return [part for part in parts if part]
+
+
+class KokoroRu:
+    """Loaded model plus the G2P front-end, behind a single inference lock."""
+
+    def __init__(self) -> None:
+        self._torch: torch | None = None
+        self._g2p = None
+        self._models: dict[str, torch.nn.Module] = {}
+        self._packs: dict[str, torch.Tensor] = {}
+        # The Albert encoder and the iSTFTNet decoder keep per-call scratch
+        # buffers; concurrent forwards on one model interleave into them. The
+        # model is fast enough on CPU that serialising is not the bottleneck.
+        self._lock = threading.Lock()
+
+    def load(self) -> None:
+        import torch
+        from kokoro import KModel
+
+        torch.set_num_threads(THREADS)
+        self._torch = torch
+
+        # RuG2P is imported from the baked snapshot, not installed, and it
+        # resolves espeak-data/ plus kokoro-config.json next to itself.
+        sys.path.insert(0, str(MODEL_DIR))
+        from ru_g2p import RuG2P
+
+        self._g2p = RuG2P(
+            espeak_data=MODEL_DIR / "espeak-data",
+            vocab_path=MODEL_DIR / "kokoro-config.json",
+        )
+
+        for stem in sorted({stem for stem, _ in VOICE_SPECS.values()}):
+            checkpoint = MODEL_DIR / f"{stem}.pth"
+            if not checkpoint.exists():
+                log.warning("checkpoint %s missing, voices using it stay unavailable", checkpoint)
+                continue
+            # repo_id is only used to build the default model filename; passing
+            # both config and model keeps it from touching the HF cache at all.
+            self._models[stem] = KModel(
+                repo_id=str(MODEL_DIR),
+                config=str(MODEL_DIR / "config.json"),
+                model=str(checkpoint),
+            ).eval()
+            log.info("loaded checkpoint %s", checkpoint.name)
+
+        for name in VOICE_SPECS:
+            pack = MODEL_DIR / "voices" / f"{name}.pt"
+            if pack.exists():
+                self._packs[name] = torch.load(str(pack), map_location="cpu", weights_only=True)
+
+        if not self.available_voices():
+            raise RuntimeError(f"no usable voices under {MODEL_DIR}")
+
+    def available_voices(self) -> list[str]:
+        return [
+            name
+            for name in VOICE_SPECS
+            if name in self._packs and VOICE_SPECS[name][0] in self._models
+        ]
+
+    def phonemes(self, text: str):
+        for chunk in split_text(text):
+            ps, _oov = self._g2p.phonemize(chunk)
+            ps = ps.strip()
+            if ps:
+                yield from split_phonemes(ps)
+
+    def synthesize(self, text: str, voice: str, speed: float) -> np.ndarray:
+        torch = self._torch
+        assert torch is not None, "synthesize() before load()"
+        stem, _gender = VOICE_SPECS[voice]
+        model = self._models[stem]
+        pack = self._packs[voice]
+
+        gap = torch.zeros(int(CHUNK_GAP_S * SAMPLE_RATE), dtype=torch.float32)
+        pieces: list[torch.Tensor] = []
+        with self._lock:
+            for ps in self.phonemes(text):
+                # The style vector is picked by phoneme-string length, which is
+                # why the model sounds deterministic for identical text.
+                style = pack[len(ps) - 1]
+                # The packs ship as [510, 256]; KModel wants a batch of one.
+                if style.dim() == 1:
+                    style = style.unsqueeze(0)
+                if pieces:
+                    pieces.append(gap)
+                pieces.append(model(ps, style, speed, return_output=True).audio)
+
+        if not pieces:
+            return np.zeros(0, dtype=np.float32)
+        return torch.cat(pieces).numpy().astype(np.float32, copy=False)
+
+
+def encode(audio: np.ndarray, fmt: str) -> bytes:
+    clipped = np.clip(audio, -1.0, 1.0)
+    if fmt == "pcm":
+        # OpenAI's pcm is raw signed 16-bit little-endian mono at 24 kHz.
+        return (clipped * 32767.0).astype("<i2").tobytes()
+
+    buffer = io.BytesIO()
+    with wave.open(buffer, "wb") as out:
+        out.setnchannels(1)
+        out.setsampwidth(2)
+        out.setframerate(SAMPLE_RATE)
+        out.writeframes((clipped * 32767.0).astype("<i2").tobytes())
+    wav = buffer.getvalue()
+
+    if fmt == "wav":
+        return wav
+
+    import imageio_ffmpeg
+
+    command = [
+        imageio_ffmpeg.get_ffmpeg_exe(),
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-i",
+        "pipe:0",
+        "-ar",
+        str(SAMPLE_RATE),
+        "-ac",
+        "1",
+        *FFMPEG_ARGS[fmt],
+        "-f",
+        FFMPEG_CONTAINERS[fmt],
+        "pipe:1",
+    ]
+    done = subprocess.run(command, input=wav, capture_output=True, check=False)
+    if done.returncode != 0:
+        raise RuntimeError(done.stderr.decode("utf-8", "replace").strip()[-400:])
+    return done.stdout
+
+
+engine = KokoroRu()
+state: dict[str, str | None] = {"status": "loading", "error": None}
+
+
+def boot() -> None:
+    try:
+        engine.load()
+        state["status"] = "ready"
+        log.info("ready: voices=%s", ", ".join(engine.available_voices()))
+    except Exception as exc:
+        state["status"] = "error"
+        state["error"] = f"{type(exc).__name__}: {exc}"
+        log.exception("model failed to load")
+
+
+@asynccontextmanager
+async def lifespan(_app: FastAPI):
+    # Off the event loop: loading pulls ~700 MB of weights and runs three ONNX
+    # sessions, and /healthz has to stay answerable while it happens.
+    threading.Thread(target=boot, name="kokoro-load", daemon=True).start()
+    yield
+
+
+app = FastAPI(title="kokoro-ru OpenAI TTS", version="1.0.0", lifespan=lifespan)
+
+Format = Literal["mp3", "opus", "aac", "flac", "wav", "pcm"]
+
+
+class SpeechRequest(BaseModel):
+    # `protected_namespaces` silences pydantic's warning about the `model_`
+    # prefix; `extra="ignore"` absorbs the fields newer OpenAI clients add
+    # (instructions, the legacy `format` alias) without failing the request.
+    model_config = ConfigDict(extra="ignore", protected_namespaces=())
+
+    input: str = Field(min_length=1)
+    model: str = MODEL_ID
+    voice: str | None = None
+    response_format: Format = "wav"
+    speed: float | None = Field(default=None, ge=0.25, le=4.0)
+
+
+def fail(status: int, message: str, param: str | None = None, code: str | None = None) -> JSONResponse:
+    return JSONResponse(
+        status_code=status,
+        content={
+            "error": {
+                "message": message,
+                "type": "invalid_request_error" if status < 500 else "server_error",
+                "param": param,
+                "code": code,
+            }
+        },
+    )
+
+
+def resolve_voice(requested: str | None) -> str | None:
+    name = (requested or DEFAULT_VOICE).strip().lower()
+    name = VOICE_ALIASES.get(name, name)
+    return name if name in engine.available_voices() else None
+
+
+# response_model=None: the handler returns a Response subclass directly, and
+# FastAPI would otherwise try to build a Pydantic model out of the union.
+@app.post("/v1/audio/speech", response_model=None)
+def create_speech(request: SpeechRequest) -> Response | JSONResponse:
+    if state["status"] != "ready":
+        return fail(503, f"model is not ready: {state['status']}", code="model_not_ready")
+
+    voice = resolve_voice(request.voice)
+    if voice is None:
+        available = ", ".join(engine.available_voices())
+        return fail(
+            400,
+            f"unknown voice {request.voice!r}; available: {available}",
+            param="voice",
+            code="unknown_voice",
+        )
+
+    try:
+        audio = engine.synthesize(request.input, voice, request.speed or 1.0)
+    except Exception as exc:
+        log.exception("synthesis failed")
+        return fail(500, f"synthesis failed: {exc}", code="synthesis_failed")
+
+    if audio.size == 0:
+        return fail(
+            400,
+            "input contains no speakable text for the Russian G2P",
+            param="input",
+            code="no_phonemes",
+        )
+
+    try:
+        payload = encode(audio, request.response_format)
+    except Exception as exc:
+        log.exception("encoding to %s failed", request.response_format)
+        return fail(500, f"encoding to {request.response_format} failed: {exc}", code="encoding_failed")
+
+    return Response(
+        content=payload,
+        media_type=CONTENT_TYPES[request.response_format],
+        headers={"model-id": MODEL_ID, "voice-id": voice},
+    )
+
+
+@app.get("/v1/models")
+def list_models() -> dict:
+    return {
+        "object": "list",
+        "data": [{"id": MODEL_ID, "object": "model", "created": MODEL_CREATED, "owned_by": "zaakirio"}],
+    }
+
+
+@app.get("/v1/voices")
+def list_voices() -> dict:
+    return {
+        "object": "list",
+        "ready": state["status"] == "ready",
+        "data": [
+            {"id": name, "object": "voice", "checkpoint": VOICE_SPECS[name][0], "gender": VOICE_SPECS[name][1]}
+            for name in engine.available_voices()
+        ],
+    }
+
+
+@app.get("/healthz")
+def healthz() -> JSONResponse:
+    ready = state["status"] == "ready"
+    return JSONResponse(
+        status_code=200 if ready else 503,
+        content={
+            "status": state["status"],
+            "model": MODEL_ID,
+            "voices": engine.available_voices(),
+            "sample_rate": SAMPLE_RATE,
+            "error": state["error"],
+        },
+    )

modules/containers/kokoro-tts/fetch_assets.py

@@ -0,0 +1,82 @@
+"""Bake every kokoro-ru asset the server needs into the image.
+
+Two things make a plain `FROM python` image useless for this model at runtime,
+and both are fixed here at build time:
+
+  * kokoro-ru's checkpoints and its recompiled espeak-ng data live in the HF
+    cache by default, and the HF cache is part of the disposable container
+    layer, so every `podman run` would re-download ~700 MB.
+  * ruaccent writes its ONNX models, dictionaries and Koziev data into its own
+    `site-packages/ruaccent` directory. It only downloads when those files are
+    missing, so a single `load()` here means the runtime never touches the
+    network.
+
+RuG2P resolves espeak-data/ and kokoro-config.json relative to ru_g2p.py, so
+the snapshot layout has to stay flat inside KOKORO_MODEL_DIR.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+REPO = os.environ.get("KOKORO_RU_REPO", "zaakirio/kokoro-ru")
+# A commit, not a branch: "main" would silently change the weights under a
+# rebuild that only touched an unrelated line of the Nix module.
+REVISION = os.environ.get("KOKORO_RU_REVISION", "main")
+DEST = Path(os.environ.get("KOKORO_MODEL_DIR", "/app/kokoro-ru"))
+
+VOICES = [v.strip() for v in os.environ.get("KOKORO_RU_VOICES", "sveta,masha,dima").split(",") if v.strip()]
+
+# sveta and masha share one checkpoint and differ only by voicepack, so the two
+# female voices cost one 327 MB download, not two.
+CHECKPOINTS = {
+    "sveta": "kokoro-ru-v2-base.pth",
+    "masha": "kokoro-ru-v2-base.pth",
+    "dima": "kokoro-ru-v2-dima.pth",
+}
+
+PATTERNS = [
+    # KModel reads config.json; RuG2P reads kokoro-config.json for the phoneme
+    # vocab. They are not the same file and both are required.
+    "config.json",
+    "kokoro-config.json",
+    "ru_g2p.py",
+    # Stock espeak-ng ru_dict ignores combining-acute stress marks, which is the
+    # one thing this whole front-end exists to fix. The model repo ships a
+    # recompiled dictsource; there is no substitute to fall back to.
+    "espeak-data/**",
+    *(CHECKPOINTS[v] for v in VOICES if v in CHECKPOINTS),
+    *(f"voices/{v}.pt" for v in VOICES),
+]
+
+
+def main() -> None:
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
+
+    DEST.mkdir(parents=True, exist_ok=True)
+    snapshot_download(
+        repo_id=REPO,
+        revision=REVISION,
+        allow_patterns=PATTERNS,
+        local_dir=str(DEST),
+    )
+    logging.info("kokoro-ru assets in %s at %s", DEST, REVISION)
+
+    missing = [name for name in VOICES if not (DEST / "voices" / f"{name}.pt").exists()]
+    if missing:
+        raise SystemExit(f"voice packs missing after download: {missing}")
+
+    # Warm ruaccent into site-packages so `load()` short-circuits at runtime.
+    from ruaccent import RUAccent
+
+    accent = RUAccent()
+    accent.load(omograph_model_size="turbo3.1", use_dictionary=True, tiny_mode=False)
+    logging.info("ruaccent warm: %s", accent.process_all("Здравствуйте, как ваши дела?"))
+
+
+if __name__ == "__main__":
+    main()

modules/containers/kokoro-tts/requirements.txt

@@ -0,0 +1,21 @@
+# torch is installed separately in the Dockerfile from the CPU-only index;
+# do not add it here or pip would pull the ~2.5 GB CUDA build over it.
+#
+# ru_g2p.py (from zaakirio/kokoro-ru) imports three things stock kokoro does not
+# pull on its own: the espeak-ng backend of misaki, ruaccent for stress, and the
+# phonemizer fork whose EspeakWrapper misaki drives.
+kokoro==0.9.4
+misaki[en]>=0.9.4
+phonemizer-fork
+espeakng-loader
+ruaccent
+
+# kokoro's Albert encoder and ruaccent's ONNX exports both go through
+# transformers. ru_g2p.py shims token_type_ids for v5, so the floor is what
+# matters, not the ceiling.
+transformers>=4.46
+
+fastapi
+uvicorn
+imageio-ffmpeg
+numpy>=1.26,<3

modules/options.nix

@@ -38,4 +38,4 @@
       '';
     };
   };
-}
+}

modules/users.nix

@@ -45,11 +45,21 @@ in
         isNormalUser = true;
         group = "users";
         # Pinned, not left to NixOS' nextfree logic: the ntfs3/exfat mount
-        # helpers (lib/xlib/helpers.nix) write the same uid into their mount
-        # options, so both sides have to agree or NTFS files show up as owned
-        # by `nobody`. NixOS has no per-user `gid` option — the primary group
-        # id comes from `group` above.
-        uid = xlib.device.uid;
+        # helpers (lib/xlib/helpers.nix) bake xlib.device.uid into their mount
+        # options, so normally both sides agree and NTFS/exFAT files do not
+        # show up as owned by `nobody`. NixOS has no per-user `gid` option —
+        # the primary group id comes from `group` above.
+        #
+        # sapphira is the one exception, and only until its filesystem gets
+        # migrated: /var/lib/nixos/uid-map still reserves 1000 for the
+        # long-removed `yuyus` and NixOS never renumbers an existing user, so
+        # the live `oqyude` there is uid 1001. Without this branch a rebuild
+        # would rewrite the user to 1000 while every file is still owned by
+        # 1001. The cost: the exFAT mounts on sapphira still get uid=1000 from
+        # xlib.device.uid, so that user cannot write to /mnt/archive or
+        # /mnt/mobile until the id question is settled.
+        # TODO: delete this branch once sapphira is migrated to 1000.
+        uid = if xlib.device.hostname == "sapphira" then 1001 else xlib.device.uid;
         description = "Jor Oqyude";
         hashedPasswordFile = config.sops.secrets.hashed_password.path; # hashed_password
         homeMode = "700";

modules/wsl/containers/default.nix

@@ -7,7 +7,7 @@
 {
   imports = [
     # shared container modules live in ../../containers
-    ../../containers/silero-tts.nix
+    ../../containers/kokoro-tts.nix
   ];
 
   environment.systemPackages = with pkgs; [