diff --git a/stacks/cosyvoice/.env.example b/stacks/cosyvoice/.env.example
new file mode 100644
index 0000000..b2984b8
--- /dev/null
+++ b/stacks/cosyvoice/.env.example
@@ -0,0 +1,42 @@
+# CosyVoice stack tunables. Copy to `.env` on irv-ml1 before deploying.
+#
+#   cp .env.example .env
+#   # edit as needed
+#   docker compose up -d
+
+# Image tag. `v1.3.2` (2026-01-18) is the latest wrapper and ships
+# Fun-CosyVoice3-0.5B-2512 (CosyVoice 3 model). The older `v3.x`
+# tag family (2025-12-18) also ships CosyVoice 3 — the wrapper
+# version numbering diverged from the model version, which is
+# confusing but deliberate upstream. Stay on v1.3.x.
+#
+# See https://hub.docker.com/r/neosun/cosyvoice/tags for updates.
+COSYVOICE_VERSION=v1.3.2
+
+# Host port for the web UI + REST API (container listens on 8188).
+# Avoiding 8188 on the host since ComfyUI already has it.
+COSYVOICE_PORT=8190
+
+# Bind address. 0.0.0.0 exposes on all interfaces including the WG
+# tunnel IP (10.100.79.3). Use 127.0.0.1 to restrict to local-only.
+COSYVOICE_BIND=0.0.0.0
+
+# Path inside the container for the active model. The image places
+# Fun-CosyVoice3-0.5B at pretrained_models/Fun-CosyVoice3-0.5B on
+# first run — keep this default unless upstream ships alternate
+# model variants.
+COSYVOICE_MODEL_DIR=pretrained_models/Fun-CosyVoice3-0.5B
+
+# Seconds of idleness before the server releases GPU VRAM. Reloading
+# takes a few seconds; trade off fast re-use vs sharing GPUs with
+# other workloads (llama-swap style). Default 600 (10 min).
+COSYVOICE_GPU_IDLE_TIMEOUT=600
+
+# Host paths for persistent data. Must exist and be writable by the
+# container before first `up`.
+#   voices/  cloned speaker profiles — precious, restic-covered
+#   input/   uploaded source audio   — scratch
+#   output/  synthesized clips       — scratch, regenerable
+COSYVOICE_VOICES_DIR=/worktank/cosyvoice/voices
+COSYVOICE_INPUT_DIR=/worktank/cosyvoice/input
+COSYVOICE_OUTPUT_DIR=/worktank/cosyvoice/output
diff --git a/stacks/cosyvoice/README.md b/stacks/cosyvoice/README.md
new file mode 100644
index 0000000..18b8426
--- /dev/null
+++ b/stacks/cosyvoice/README.md
@@ -0,0 +1,104 @@
+# CosyVoice
+
+Multilingual expressive TTS with zero-shot voice cloning, served by
+the `neosun/cosyvoice` wrapper around FunAudioLLM's
+Fun-CosyVoice3-0.5B-2512.
+
+**Server:** irv-ml1 (Irvine, WireGuard-only)
+**Port:** 8190 (configurable via `.env`; container listens on 8188
+internally but host port moved to avoid ComfyUI's 8188)
+**GPUs:** both exposed (`NVIDIA_VISIBLE_DEVICES=all`); image reads
+`CUDA_VISIBLE_DEVICES` for pinning
+**Upstream wrapper:** [neosun100/cosyvoice-docker](https://github.com/neosun100/cosyvoice-docker)
+**Upstream model:** [FunAudioLLM/Fun-CosyVoice3-0.5B-2512](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512)
+
+## Why CosyVoice 3 (and not Kokoro / v2)
+
+Emotive content was the deal-breaker for Kokoro. CosyVoice 3 extended
+the v2 instruction-following dataset from 1,500 → 5,000 hours
+specifically covering emotions, speed, tones, dialects, accents, and
+role-playing. Streaming TTFB stays at ~150 ms.
+
+Two ways to request emotion / style:
+
+- **XML tags** — `<angry>That's my line!</angry>`, `<sad>…</sad>`,
+  `<fast>…</fast>`, `<slow>…</slow>`, `<peppa>…</peppa>`,
+  `<robot>…</robot>`
+- **Instruction prompts** — `You are a helpful assistant. 请用尽可能快地语速说一句话。<|endofprompt|>`
+  gives finer control via natural-language directives in the
+  instruction channel
+
+Language center of gravity is Mandarin + Cantonese (18+ Chinese
+dialects) and then 8 other languages (English, Japanese, Korean,
+German, Spanish, French, Italian, Russian). English works fine but
+don't expect ElevenLabs-grade English prosody polish — ear-test with
+your actual content.
+
+## API endpoints
+
+| Method + path | Purpose |
+|---|---|
+| `POST /v1/audio/speech` | OpenAI drop-in for TTS |
+| `POST /v1/voices/create` | Clone a speaker from reference audio (auto-transcription via built-in ASR) |
+| `GET /v1/voices` | List cloned voices by `voice_id` |
+| `GET /health` | Health probe (used by docker healthcheck) |
+
+## Path layout
+
+| Host path | Container path | Purpose | Restic? |
+|---|---|---|---|
+| `/worktank/cosyvoice/voices/` | `/data/voices` | Cloned speaker profiles | **included** (precious — reproducing a clone needs the original reference audio) |
+| `/worktank/cosyvoice/input/` | `/data/input` | Scratch for uploaded source audio | excluded |
+| `/worktank/cosyvoice/output/` | `/data/output` | Synthesized clips | excluded (regenerable) |
+
+**Model weights (~2–3 GB) are NOT bind-mounted.** The image places
+them at `pretrained_models/Fun-CosyVoice3-0.5B/` inside the container
+on first run. Docker's image-layer cache keeps them across normal
+`compose up -d` recreates; a `docker image rm` or tag bump re-downloads.
+
+## First-time deploy on irv-ml1
+
+```bash
+# 1. Push compose + env template
+scripts/deploy-stack.sh irv-ml1 cosyvoice
+
+# 2. Create the host dirs. One-time sudo — /worktank is root-owned.
+ssh -t irv-ml1 'sudo mkdir -p /worktank/cosyvoice/{voices,input,output} && \
+                sudo chown -R lkraven:lkraven /worktank/cosyvoice'
+
+# 3. Pull + up. First `up` downloads ~2–3 GB of model weights inside
+#    the container; allow a few minutes before the health probe passes.
+ssh irv-ml1 '
+  cd /opt/docker/compose/cosyvoice && \
+  cp -n .env.example .env && \
+  docker compose config >/dev/null && \
+  docker compose pull && \
+  docker compose up -d && \
+  docker compose logs -f --tail=30
+'
+```
+
+## Smoke test
+
+```bash
+# From the workstation over WG — OpenAI-compatible request
+curl -X POST http://10.100.79.3:8190/v1/audio/speech \
+  -H 'Content-Type: application/json' \
+  -d '{"model":"cosyvoice","voice":"default","input":"<angry>Hello there.</angry>","response_format":"wav"}' \
+  -o /tmp/out.wav
+```
+
+## Version bump
+
+```bash
+# Pick a new tag from https://hub.docker.com/r/neosun/cosyvoice/tags
+ssh irv-ml1 '
+  cd /opt/docker/compose/cosyvoice && \
+  sed -i "s/^COSYVOICE_VERSION=.*/COSYVOICE_VERSION=<new-tag>/" .env && \
+  docker compose pull && \
+  docker compose up -d
+'
+```
+
+Voices / input / output persist across bumps. Model weights inside
+the image re-download on first run of the new tag.
diff --git a/stacks/cosyvoice/compose.yaml b/stacks/cosyvoice/compose.yaml
new file mode 100644
index 0000000..28d89dd
--- /dev/null
+++ b/stacks/cosyvoice/compose.yaml
@@ -0,0 +1,60 @@
+# CosyVoice — multilingual expressive TTS with voice cloning.
+#
+# Ships the Fun-CosyVoice3-0.5B-2512 model from FunAudioLLM (latest
+# as of 2026-04). Streaming PCM chunks with ~150 ms TTFB. Emotional
+# control via either XML tags (<angry>text</angry>) or instruction
+# prompts (`You are a helpful assistant. <|endofprompt|>` syntax).
+#
+# Runs on irv-ml1 (dual GPU). Both GPUs exposed via
+# NVIDIA_VISIBLE_DEVICES=all; image reads CUDA_VISIBLE_DEVICES if
+# you later want to pin.
+#
+# Path split:
+#   /worktank/cosyvoice/voices  → /data/voices   cloned speaker profiles
+#   /worktank/cosyvoice/input   → /data/input    scratch for uploaded source audio
+#   /worktank/cosyvoice/output  → /data/output   synthesized clips
+#
+# Models (~2–3 GB) download on first run into the image's internal
+# pretrained_models/ path. Not bind-mounted (the image expects an
+# exact layout we don't have authoritative docs for); recreating the
+# container without the cached image will re-download. Cached image
+# layer persists through `compose up -d` recreates.
+#
+# API routes (OpenAI-compatible where marked):
+#   POST /v1/audio/speech         — OpenAI drop-in for TTS
+#   POST /v1/voices/create        — voice cloning (reference audio in)
+#   GET  /v1/voices               — list cloned voices
+#   GET  /health                  — health probe
+#
+# All tunables live in .env — edit that, not this file.
+
+services:
+  cosyvoice:
+    image: neosun/cosyvoice:${COSYVOICE_VERSION}
+    container_name: cosyvoice
+    restart: unless-stopped
+    runtime: nvidia
+    ports:
+      - "${COSYVOICE_BIND:-0.0.0.0}:${COSYVOICE_PORT}:8188"
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - MODEL_DIR=${COSYVOICE_MODEL_DIR:-pretrained_models/Fun-CosyVoice3-0.5B}
+      - PORT=8188
+      - GPU_IDLE_TIMEOUT=${COSYVOICE_GPU_IDLE_TIMEOUT:-600}
+    volumes:
+      - ${COSYVOICE_VOICES_DIR}:/data/voices
+      - ${COSYVOICE_INPUT_DIR}:/data/input
+      - ${COSYVOICE_OUTPUT_DIR}:/data/output
+    healthcheck:
+      test: ["CMD-SHELL", "curl -fsS http://localhost:8188/health >/dev/null || exit 1"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      # First boot pulls ~2–3 GB of model weights.
+      start_period: 300s
+    labels:
+      - homepage.group=AI Systems
+      - homepage.name=CosyVoice
+      - homepage.icon=mdi-account-voice
+      - homepage.description=Expressive multilingual TTS + cloning (irv-ml1)
+      - homepage.href=http://10.100.79.3:${COSYVOICE_PORT}
diff --git a/stacks/parakeet/.env.example b/stacks/parakeet/.env.example
new file mode 100644
index 0000000..8319035
--- /dev/null
+++ b/stacks/parakeet/.env.example
@@ -0,0 +1,49 @@
+# Parakeet ASR stack tunables. Copy to `.env` on irv-ml1 before deploying.
+#
+#   cp .env.example .env
+#   # edit as needed
+#   docker compose build
+#   docker compose up -d
+
+# Pinned git SHA to build from. Bump + rebuild when you want upstream
+# fixes. `main` latest as of 2026-04:
+#   https://github.com/Shadowfita/parakeet-tdt-0.6b-v2-fastapi
+PARAKEET_SHA=31c5652b62d09653ad5ea8190c0ad0d35394174d
+
+# Host port for the FastAPI server (container listens on 8000)
+PARAKEET_PORT=8765
+
+# Bind address. 0.0.0.0 exposes on all interfaces including the WG
+# tunnel IP (10.100.79.3). Use 127.0.0.1 to restrict to local-only.
+PARAKEET_BIND=0.0.0.0
+
+# Host path for the HuggingFace cache (parakeet-tdt-0.6b-v2 weights
+# ~2.5 GB). Persistent across container recreates. Must exist before
+# first `up` with ownership matching the container user (root inside
+# this image — no UID juggling needed, but the host dir needs to be
+# writable by the container).
+PARAKEET_MODELS_DIR=/worktank/parakeet/models
+
+# Inference precision. fp16 halves VRAM and is lossless for parakeet
+# in practice; use fp32 only if fp16 shows degraded WER for your
+# domain audio.
+PARAKEET_MODEL_PRECISION=fp16
+
+# Batch size for the transcribe queue. Larger = better throughput
+# under load at the cost of per-request latency.
+PARAKEET_BATCH_SIZE=4
+
+# Max single-clip duration (seconds). Longer inputs get rejected
+# by the server with 400. Upstream default.
+PARAKEET_MAX_AUDIO_DURATION=30
+
+# Silero VAD threshold (0–1). Higher = stricter about what counts
+# as speech (fewer false wake-ups on silence, more chance of clipping
+# soft speech). 0.5 is upstream default.
+PARAKEET_VAD_THRESHOLD=0.5
+
+# End-to-end processing timeout per request (seconds).
+PARAKEET_PROCESSING_TIMEOUT=60
+
+# Log level: DEBUG | INFO | WARNING | ERROR
+PARAKEET_LOG_LEVEL=INFO
diff --git a/stacks/parakeet/README.md b/stacks/parakeet/README.md
new file mode 100644
index 0000000..0cbe4da
--- /dev/null
+++ b/stacks/parakeet/README.md
@@ -0,0 +1,88 @@
+# Parakeet ASR
+
+NVIDIA Parakeet-TDT 0.6B v2 speech-to-text served via a FastAPI
+wrapper with Silero VAD and WebSocket streaming.
+
+**Server:** irv-ml1 (Irvine, WireGuard-only)
+**Port:** 8765 (configurable via `.env`)
+**GPUs:** both exposed (`NVIDIA_VISIBLE_DEVICES=all`); upstream
+respects `CUDA_VISIBLE_DEVICES` if later pinning is needed
+**Upstream:** [Shadowfita/parakeet-tdt-0.6b-v2-fastapi](https://github.com/Shadowfita/parakeet-tdt-0.6b-v2-fastapi)
+**Image:** built locally from a pinned git SHA via docker buildx's
+git URL context — no source vendored into this workspace
+
+## API endpoints
+
+| Method + path | Purpose |
+|---|---|
+| `POST /transcribe` | Batch transcription (multipart file upload) |
+| `WS /ws/transcribe` | Streaming with Silero VAD — partial + final segments |
+| `GET /healthz` | Health probe (used by docker healthcheck) |
+
+Not the literal OpenAI `/v1/audio/transcriptions` path. If you have
+a downstream client that demands that URL shape, either point its
+base URL at `/transcribe`, or add a Traefik/nginx path alias in
+front.
+
+## Path layout
+
+| Host path | Container path | Purpose | Restic? |
+|---|---|---|---|
+| `/worktank/parakeet/models/` | `/models` (`HF_HOME`) | HF cache for parakeet-tdt-0.6b-v2 weights (~2.5 GB) | excluded (regenerable from HF) |
+
+The container runs as root internally; host dir just needs to exist
+and be writable.
+
+## First-time deploy on irv-ml1
+
+```bash
+# 1. Push compose + env template
+scripts/deploy-stack.sh irv-ml1 parakeet
+
+# 2. Create the models dir on the host. One-time sudo — /worktank
+#    itself is root-owned.
+ssh -t irv-ml1 'sudo mkdir -p /worktank/parakeet/models && \
+                sudo chown -R lkraven:lkraven /worktank/parakeet'
+
+# 3. Build the image (first time only; ~5–10 min for torch + NeMo
+#    wheels). Then `up`.
+ssh irv-ml1 '
+  cd /opt/docker/compose/parakeet && \
+  cp -n .env.example .env && \
+  docker compose config >/dev/null && \
+  docker compose build && \
+  docker compose up -d && \
+  docker compose logs -f --tail=30
+'
+```
+
+First `/transcribe` request downloads parakeet-tdt-0.6b-v2 weights
+to `/worktank/parakeet/models/` (~2.5 GB).
+
+## Smoke test
+
+```bash
+# From the workstation over WG
+curl -F "file=@sample.wav" http://10.100.79.3:8765/transcribe
+```
+
+## Rebuild against a newer upstream commit
+
+```bash
+ssh irv-ml1 '
+  cd /opt/docker/compose/parakeet && \
+  sed -i "s/^PARAKEET_SHA=.*/PARAKEET_SHA=<new-sha>/" .env && \
+  docker compose build && \
+  docker compose up -d
+'
+```
+
+Models dir is unaffected.
+
+## Deploy updates
+
+```bash
+# After editing compose.yaml or .env.example here
+scripts/deploy-stack.sh irv-ml1 parakeet
+ssh irv-ml1 'cd /opt/docker/compose/parakeet && docker compose up -d'
+```
diff --git a/stacks/parakeet/compose.yaml b/stacks/parakeet/compose.yaml
new file mode 100644
index 0000000..9b35fb5
--- /dev/null
+++ b/stacks/parakeet/compose.yaml
@@ -0,0 +1,67 @@
+# Parakeet ASR — NVIDIA Parakeet-TDT 0.6B v2 speech-to-text.
+#
+# Wraps Shadowfita/parakeet-tdt-0.6b-v2-fastapi (FastAPI + Silero VAD +
+# WebSocket streaming). Upstream provides no prebuilt image, so we
+# build from a pinned git commit via docker buildx's git URL context
+# — no source files vendored into this repo.
+#
+# Runs on irv-ml1 (dual GPU). Both GPUs exposed; upstream respects
+# CUDA_VISIBLE_DEVICES if you want to pin later.
+#
+# HF weights (~2.5 GB for parakeet-tdt-0.6b-v2) cache to
+# ${PARAKEET_MODELS_DIR} via HF_HOME=/models, persistent across
+# container recreates.
+#
+# API routes (per upstream README):
+#   POST /transcribe          — batch transcription
+#   WS   /ws/transcribe       — streaming with Silero VAD
+#   GET  /healthz
+#
+# Not a literal OpenAI `/v1/audio/transcriptions` path; point clients
+# at /transcribe directly, or add a reverse-proxy alias if drop-in
+# compat is needed later.
+#
+# First `up` triggers a fresh docker build (python:3.10-slim +
+# torch + NeMo ≈ 5–10 min). Subsequent starts reuse the cached
+# image unless PARAKEET_SHA changes.
+#
+# All tunables live in .env — edit that, not this file.
+
+services:
+  parakeet:
+    image: local/parakeet:${PARAKEET_SHA}
+    build:
+      context: https://github.com/Shadowfita/parakeet-tdt-0.6b-v2-fastapi.git#${PARAKEET_SHA}
+    container_name: parakeet
+    restart: unless-stopped
+    runtime: nvidia
+    ports:
+      - "${PARAKEET_BIND:-0.0.0.0}:${PARAKEET_PORT}:8000"
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - HF_HOME=/models
+      - DEVICE=cuda
+      - MODEL_PRECISION=${PARAKEET_MODEL_PRECISION:-fp16}
+      - BATCH_SIZE=${PARAKEET_BATCH_SIZE:-4}
+      - TARGET_SAMPLE_RATE=16000
+      - MAX_AUDIO_DURATION=${PARAKEET_MAX_AUDIO_DURATION:-30}
+      - VAD_THRESHOLD=${PARAKEET_VAD_THRESHOLD:-0.5}
+      - PROCESSING_TIMEOUT=${PARAKEET_PROCESSING_TIMEOUT:-60}
+      - LOG_LEVEL=${PARAKEET_LOG_LEVEL:-INFO}
+    volumes:
+      - ${PARAKEET_MODELS_DIR}:/models
+    healthcheck:
+      test: ["CMD-SHELL", "curl -fsS http://localhost:8000/healthz >/dev/null || exit 1"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      # First `up` may spend several minutes on torch/NeMo install
+      # during the image build phase; after the image exists, startup
+      # is ~30-60s (NeMo model load).
+      start_period: 180s
+    labels:
+      - homepage.group=AI Systems
+      - homepage.name=Parakeet ASR
+      - homepage.icon=mdi-microphone
+      - homepage.description=Parakeet-TDT speech-to-text (irv-ml1)
+      - homepage.href=http://10.100.79.3:${PARAKEET_PORT}