GPU TTS Docker Launcher

Standalone Windows PowerShell launcher for local, Dockerized CUDA TTS voice-clone and preset-voice testing with inference-speed comparison.

PowerShell acts as the frontend menu. Docker runs one selected model container at a time, writes generated WAV/JSON files to output, and removes the selected one-off container when you quit so the model unloads from GPU memory.

What This Is For

This repo is for local CUDA TTS testing: load one model at a time, synthesize text against a local reference voice or configured preset voice, and compare how fast inference runs on your own GPU.

Included model containers:

• xtts: XTTS-v2 voice clone. No official female preset catalog is bundled.
• qwen: Qwen3-TTS 12Hz 0.6B Base voice clone, plus official Qwen CustomVoice female presets.
• miotts: MioTTS 0.1B voice clone, plus the official MioTTS English female preset embedding.
• luxtts: LuxTTS 100M voice clone. No official female preset catalog is bundled.
• piper: Piper TTS preset voice picker with American female, British female, and all high-quality upstream preset options. Piper does not do zero-shot voice cloning.
• kokoro: Kokoro-82M preset voice picker with all American female and British female presets from the model. Kokoro does not do zero-shot voice cloning.
• kitten: KittenTTS Nano 15M, Micro 40M, or Mini 80M ONNX model picker with all female presets. KittenTTS does not do zero-shot voice cloning.
• supertonic: Supertonic 3 99M multilingual ONNX model with all five official female preset styles. The open-weight package does not perform local reference-WAV cloning.
• chatterbox: Resemble AI Chatterbox 500M voice clone, plus the official built-in default voice fallback.
• sopro: SoproTTS 135M English voice clone. No official female preset catalog is bundled.
• pockettts: Kyutai PocketTTS 100M voice clone, plus selected official English built-in voices. PocketTTS is CPU-first upstream.
• moss: OpenMOSS MOSS-TTS-Nano 100M ONNX voice clone, plus the official English Female built-in voices Ava and Bella.
• f5tts: SWivid F5-TTS reference voice clone. No official American/British female preset catalog is bundled.

The repo does not include voice samples, transcripts, model weights, generated outputs, API keys, or Docker image exports.

This is a local testing bench, not a production API service or hosted TTS product.

License

This project is source-available for local testing and evaluation only. It is not open-source licensed.

You may clone, modify, build, and run it locally for testing. You may not use it commercially, provide it as a hosted service, or redistribute built Docker images, containers, model weights, downloaded caches, generated service bundles, or other built artifacts.

See LICENSE for the full terms. Third-party models and dependencies keep their own licenses and usage restrictions.

Hardware Target

The default build targets NVIDIA RTX 50-series / Blackwell GPUs using PyTorch CUDA 13 wheels. The primary tested target is an RTX 5070.

Default CUDA stack:

torch==2.11.0
torchaudio==2.11.0
https://download.pytorch.org/whl/cu130

The launcher detects an RTX 5070 with nvidia-smi and exports TTS_GPU so Docker exposes only that GPU to the selected container. Inside each container, CUDA_VISIBLE_DEVICES=0 makes the selected GPU appear as device 0.

For another compatible NVIDIA GPU, pass a Docker GPU device id or UUID:

.\run.ps1 -Gpu 0

To force CUDA 12.8 wheels instead of CUDA 13:

.\run.ps1 -Cuda128

RTX 5070 Test Notes

These are informal local measurements from short voice-clone prompts on an RTX 5070. They are included only as rough comparison points for local CUDA inference speed; your results will vary with prompt length, model cache warmth, Docker/image state, driver stack, and reference audio.

No full machine details are published here.

First Run

1. Add your local voice reference files. These are intentionally ignored by git.

Recommended filenames:

reference.wav
reference.txt

reference.wav should be a clean sample of the target voice. reference.txt should contain the exact words spoken in that WAV. F5-TTS uses both files and is a good fit for American or British female reference samples. Chatterbox, SoproTTS, PocketTTS, and MOSS-TTS-Nano use the WAV but do not need the transcript. Piper, Kokoro, KittenTTS, and Supertonic 3 do not use these files because they use configured preset voices instead of zero-shot cloning.

Private voice-clone inputs are runtime inputs only. Do not put proprietary voice samples, transcripts, API keys, or per-user clone settings in Python files or commits. If another app stores the selected clone settings in a database, pass those resolved local file paths into this launcher at runtime with -Speaker and -ReferenceText, or through the model-specific environment variables.

2. Open PowerShell in this folder.

3. Run:

.\run.ps1

The menu asks which model to run. For clone-capable models, the first voice option is always your local reference voice clone. If the upstream model publishes usable female presets, those appear after the clone option. Piper, Kokoro, KittenTTS, and Supertonic 3 ask which preset voice to load. KittenTTS first asks whether to use Nano 15M, Micro 40M, or Mini 80M. After the selected model loads, type text and press Enter.

Do not type synthesis text while a model is still downloading or loading. Wait for the model-specific READY line and prompt:

xttsv2>
qwen0.6btts>
miotts0.1b>
luxtts100m>
piper>
kokoro82m>
kittentts>
supertonic3>
chatterbox>
soprotts>
pockettts>
mossttsnano>
f5tts>

Outputs are written locally and ignored by git:

output/
output/qwen/
output/miotts/
output/luxtts/
output/piper/
output/kokoro/
output/kitten/
output/supertonic/
output/chatterbox/
output/sopro/
output/pockettts/
output/moss/
output/f5tts/

Model downloads are cached under .cache and are also ignored by git.

Interactive Commands

Paste multiple lines:

:paste

Finish pasted text with a line containing only:

:end

Synthesize a mounted prompt file:

:file /workspace/prompt.txt

Quit and unload the selected container:

:q

Model Tuning

LuxTTS defaults are tuned for slower, less-skippy output:

LUXTTS_SPEED=0.68
LUXTTS_NUM_STEPS=6
LUXTTS_T_SHIFT=0.35
LUXTTS_PROMPT_DURATION=1000.0

LUXTTS_PROMPT_DURATION=1000.0 tells LuxTTS to use the full reference WAV with the full reference transcript.

MioTTS defaults are tuned to reduce drawn-out vowel/syllable loops:

MIOTTS_MAX_NEW_TOKENS=520
MIOTTS_MAX_TOKEN_OVERHEAD=80
MIOTTS_MAX_TOKENS_PER_CHAR=3.0
MIOTTS_NO_REPEAT_NGRAM_SIZE=6
MIOTTS_REPETITION_PENALTY=1.15
MIOTTS_TEMPERATURE=0.55
MIOTTS_TOP_P=0.85

Qwen includes the official CUDA 13 flash-attn-3 wheel plus a compatibility shim for Qwen's older flash_attn.flash_attn_interface import. Keep Qwen's main attention mode at sdpa; flash_attention_3 was tested on the RTX 5070 target and failed during generation with a CUDA kernel-image error.

For Qwen, the picker now offers:

Qwen's published CustomVoice list does not include an American female or British female preset. The launcher includes the official female presets only and keeps reference-clone as option 1 for your own American/British reference voice.

.\run.ps1 -Model qwen -QwenVoice reference-clone
.\run.ps1 -Model qwen -QwenVoice serena

For MioTTS, the picker now offers:

.\run.ps1 -Model miotts -MioVoice reference-clone
.\run.ps1 -Model miotts -MioVoice en-female

Chatterbox uses Resemble AI's Chatterbox 500M TTS model. The picker offers voice cloning first. Chatterbox does not publish named American female or British female presets in the official package, so the only non-clone option exposed here is the built-in default conditioning from the model.

.\run.ps1 -Model chatterbox -ChatterboxVoice reference-clone
.\run.ps1 -Model chatterbox -ChatterboxVoice builtin-default

Chatterbox defaults:

CHATTERBOX_CHUNK_CHARS=220
CHATTERBOX_PAUSE_MS=220
CHATTERBOX_EXAGGERATION=0.45
CHATTERBOX_CFG_WEIGHT=0.35
CHATTERBOX_TEMPERATURE=0.8
CHATTERBOX_TOP_P=1.0
CHATTERBOX_MIN_P=0.05
CHATTERBOX_REPETITION_PENALTY=1.2

The Chatterbox splitter preserves pasted dialogue lines as separate turns and splits long sentences around clauses before falling back to word splits.

SoproTTS uses Samuel Vitorino's Sopro 135M English TTS model. The picker offers voice cloning only. SoproTTS does not publish named American female or British female presets in the official package, and its API requires a reference audio path or prepared reference tokens.

.\run.ps1 -Model sopro -SoproVoice reference-clone

SoproTTS defaults:

SOPRO_MODEL_ID=samuel-vitorino/sopro
SOPRO_CHUNK_CHARS=220
SOPRO_PAUSE_MS=220
SOPRO_MAX_FRAMES=400
SOPRO_REF_SECONDS=12.0
SOPRO_STYLE_STRENGTH=1.2
SOPRO_TEMPERATURE=1.05
SOPRO_TOP_P=0.9

The SoproTTS splitter preserves pasted dialogue lines as separate turns and splits long sentences around clauses before falling back to word splits.

PocketTTS uses Kyutai's 100M-parameter PocketTTS model. The picker offers voice cloning first, followed by selected official English built-in voices whose names read as likely female voices. Upstream does not label these voices as American, British, female, breathy, sultry, or lower-pitched, so the catalog does not claim those attributes.

PocketTTS is intentionally CPU-first. Kyutai's README says the model runs on CPU, does not require a GPU PyTorch build, and did not show a GPU speedup in their tests. This Docker profile therefore does not reserve the RTX 5070 even though it uses the same PowerShell frontend and output workflow as the CUDA profiles.

Voice cloning requires access to Kyutai's gated PocketTTS voice-cloning weights. Accept the terms on the Hugging Face model page first. After that, either log in locally, set HF_TOKEN in the PowerShell session before launching, or paste a token into the launcher's hidden prompt when clone mode asks for it. Use a read token. If the token is fine-grained, make sure it has read access to the kyutai/pocket-tts model repo specifically. The launcher will read the token from the current environment, the normal local Hugging Face login file, or the one-time prompt and pass it into Docker for that launcher session. Do not write the token into repo files.

The launcher overrides PocketTTS clone mode to use the current gated checkpoint path hf://kyutai/pocket-tts/tts_b6369a24.safetensors. Override POCKETTTS_CLONE_WEIGHTS_PATH only if Kyutai changes the file layout again.

uvx hf auth login

# or:
$env:HF_TOKEN = "your-local-hugging-face-token"
.\run.ps1 -Model pockettts -PocketVoice reference-clone

.\run.ps1 -Model pockettts -PocketVoice reference-clone
.\run.ps1 -Model pockettts -PocketVoice alba

PocketTTS defaults:

POCKETTTS_LANGUAGE=english
POCKETTTS_CHUNK_CHARS=220
POCKETTTS_PAUSE_MS=220
POCKETTTS_MAX_TOKENS=320
POCKETTTS_FRAMES_AFTER_EOS=24
POCKETTTS_TRUNCATE_REF=1

The PocketTTS splitter preserves pasted dialogue lines as separate turns and splits long sentences around clauses before falling back to word splits.

MOSS-TTS-Nano uses OpenMOSS's 100M-parameter ONNX runtime, not the larger MOSS-TTS 8B stack. The picker offers voice cloning first, followed by the official built-in voices labeled English Female in the MOSS-TTS-Nano manifest. Upstream labels these as English Female, but does not label them American, British, breathy, sultry, or lower-pitched.

The MOSS-Nano path uses ONNX Runtime. This Docker profile defaults to MOSS_EXECUTION_PROVIDER=cuda for the RTX 5070 CUDA test bench and installs the matching CUDA PyTorch libraries so ONNX Runtime can load CUDA 13 dependencies. Set MOSS_EXECUTION_PROVIDER=cpu if the CUDA provider is not stable on your driver stack.

.\run.ps1 -Model moss -MossVoice reference-clone
.\run.ps1 -Model moss -MossVoice ava

$env:MOSS_EXECUTION_PROVIDER = "cpu"
.\run.ps1 -Model moss -MossVoice bella

MOSS-TTS-Nano defaults:

MOSS_EXECUTION_PROVIDER=cuda
MOSS_CPU_THREADS=4
MOSS_MAX_NEW_FRAMES=375
MOSS_VOICE_CLONE_MAX_TEXT_TOKENS=75
MOSS_SAMPLE_MODE=fixed
MOSS_REALTIME_STREAMING_DECODE=1

MOSS-TTS-Nano performs its own token-budget chunking during synthesis. The wrapper preloads the selected reference or built-in voice before showing READY, then writes WAV and JSON files under output/moss.

F5-TTS uses the official f5-tts package and the F5TTS_v1_Base model by default. It supports reference voice cloning from a local WAV plus the exact transcript. The upstream package does not publish a named American/British female preset catalog, so this launcher exposes reference clone first and only for F5-TTS.

.\run.ps1 -Model f5tts -F5Voice reference-clone

F5-TTS defaults:

F5TTS_MODEL_NAME=F5TTS_v1_Base
F5TTS_CHUNK_CHARS=260
F5TTS_PAUSE_MS=180
F5TTS_NFE_STEP=32
F5TTS_CFG_STRENGTH=2.0
F5TTS_SPEED=1.0
F5TTS_SWAY_SAMPLING_COEF=-1.0

The wrapper preserves pasted dialogue lines as separate turns, splits long sentences around clauses, and writes WAV/JSON files under output/f5tts.

Piper uses the public rhasspy/piper-voices ONNX voice set. The PowerShell launcher includes the original American/British female picker plus every official high quality voice at the pinned v1.0.0 model revision.

All high-quality Piper voices exposed by the picker:

You can also skip the picker and launch a known catalog id:

.\run.ps1 -Model piper -PiperVoice gb-cori
.\run.ps1 -Model piper -PiperVoice us-ljspeech
.\run.ps1 -Model piper -PiperVoice de-thorsten

The catalog lives in piper_service/voice_catalog.json. Advanced users can still override the exact files with PIPER_MODEL_FILE, PIPER_CONFIG_FILE, and PIPER_VOICE_NAME.

Piper defaults are tuned for dialogue-style pacing instead of one long fast line:

PIPER_CHUNK_CHARS=160
PIPER_PAUSE_MS=260
PIPER_SENTENCE_SILENCE=0.22
PIPER_LENGTH_SCALE=1.08

PIPER_CHUNK_CHARS controls the maximum text size for each synthesized WAV part. The Piper splitter preserves pasted dialogue lines as separate turns and splits long sentences around clauses before falling back to word splits.

Piper is wired for ONNX Runtime CUDA. The CUDA 13 path uses the current ONNX Runtime CUDA 13 nightly feed plus the PyTorch CUDA stack so it can be tested on the RTX 5070 target. If you need the CUDA 12.8 path, run the launcher with -Cuda128.

Kokoro-82M uses the official hexgrad/Kokoro-82M voice presets. The launcher includes all American female and British female presets listed for the model:

You can skip the Kokoro picker and launch a known voice id:

.\run.ps1 -Model kokoro -KokoroVoice af_heart
.\run.ps1 -Model kokoro -KokoroVoice bf_emma

Kokoro defaults are tuned to avoid rushing on longer input:

KOKORO_CHUNK_CHARS=220
KOKORO_PAUSE_MS=220
KOKORO_SPEED=0.95

The Kokoro splitter preserves pasted dialogue lines as separate turns and splits long sentences around clauses before falling back to word splits.

KittenTTS uses the official 0.8 ONNX checkpoints. After choosing KittenTTS, the launcher asks which model size to load:

Nano is the default for direct Docker runs and the first PowerShell choice. All three checkpoints expose the same female presets:

You can skip both KittenTTS pickers by supplying a size and voice id:

.\run.ps1 -Model kitten -KittenSize nano -KittenVoice bella
.\run.ps1 -Model kitten -KittenSize micro -KittenVoice luna
.\run.ps1 -Model kitten -KittenSize mini -KittenVoice rosie

KittenTTS defaults:

KITTEN_MODEL_ID=KittenML/kitten-tts-nano-0.8
KITTEN_MODEL_NAME=Nano 15M
KITTEN_CHUNK_CHARS=220
KITTEN_PAUSE_MS=220
KITTEN_SPEED=1.0

KittenTTS is wired for ONNX Runtime CUDA with the same CUDA 13 path used by Piper. If you need the CUDA 12.8 path, run the launcher with -Cuda128.

Supertonic 3 uses the official Supertone/supertonic-3 99M ONNX assets and the pinned supertonic==1.3.1 Python package. The launcher exposes the complete official female preset set:

Upstream identifies these styles only as F1 through F5. This repo does not invent American, British, breathy, sultry, pitch, or quality labels for them. The open-weight package uses fixed styles and does not turn the local reference WAV into a clone. Supertone's separate Voice Builder export workflow is not integrated.

Supertonic 3 supports 31 languages. English is the PowerShell default, and a different supported ISO language code can be passed directly:

.\run.ps1 -Model supertonic -SupertonicVoice f1
.\run.ps1 -Model supertonic -SupertonicVoice f3 -SupertonicLanguage en

Supertonic defaults:

SUPERTONIC_MODEL_NAME=supertonic-3
SUPERTONIC_LANGUAGE=en
SUPERTONIC_CHUNK_CHARS=220
SUPERTONIC_PAUSE_MS=220
SUPERTONIC_SPEED=1.0
SUPERTONIC_TOTAL_STEPS=8

The upstream Python SDK currently defaults its ONNX sessions to CPU. This Docker wrapper explicitly selects CUDAExecutionProvider and exits during startup if any of Supertonic's four ONNX sessions lacks CUDA, preventing an unnoticed CPU fallback. A local hold-open observation added roughly 776 MiB to RTX 5070 memory use, but Docker Desktop reported per-process GPU memory as N/A and other running GPU workloads can affect the total. Treat that number as a rough local observation, not a hard VRAM guarantee.

The Supertonic source package is MIT licensed and the Supertone/supertonic-3 model weights use OpenRAIL-M. Review the upstream model license before distributing outputs or using the model outside local evaluation.

Useful Options

.\run.ps1 -Model xtts
.\run.ps1 -Model qwen
.\run.ps1 -Model qwen -QwenVoice serena
.\run.ps1 -Model miotts
.\run.ps1 -Model miotts -MioVoice en-female
.\run.ps1 -Model luxtts
.\run.ps1 -Model piper
.\run.ps1 -Model piper -PiperVoice gb-cori
.\run.ps1 -Model kokoro
.\run.ps1 -Model kokoro -KokoroVoice af_heart
.\run.ps1 -Model kitten
.\run.ps1 -Model kitten -KittenSize nano -KittenVoice bella
.\run.ps1 -Model supertonic
.\run.ps1 -Model supertonic -SupertonicVoice f1
.\run.ps1 -Model chatterbox
.\run.ps1 -Model chatterbox -ChatterboxVoice builtin-default
.\run.ps1 -Model sopro
.\run.ps1 -Model sopro -SoproVoice reference-clone
.\run.ps1 -Model pockettts
.\run.ps1 -Model pockettts -PocketVoice alba
.\run.ps1 -Model moss
.\run.ps1 -Model moss -MossVoice ava
.\run.ps1 -Model f5tts
.\run.ps1 -Model f5tts -F5Voice reference-clone
.\run.ps1 -Speaker "reference.wav" -ReferenceText "reference.txt" -Language en
.\run.ps1 -NoBuild
.\run.ps1 -Model pockettts -Rebuild

Speaker files can live in this folder or in voices. This folder is mounted read-only at /workspace, and voices is mounted at /app/voices.

Do not use docker compose up --build for normal use. The model services are behind Compose profiles to prevent all models from starting together. Use .\run.ps1 as the frontend.