Skip to content

Local Models

RoomKit knows the files of the local models its providers run: sherpa-onnx speech-to-text, text-to-speech, VAD and denoiser, and Smart Turn. Each model is pinned to its URLs and their SHA-256. Download the ones you use once, while online; from then on a provider config is built by the model's name and needs no network.

python -m roomkit.local_models list
python -m roomkit.local_models download zipformer-en-2023-06-26 kokoro-int8-multi-lang-v1_0 ten-vad smart-turn-v3.2-cpu
from roomkit.voice.pipeline import (
    AudioPipelineConfig,
    SherpaOnnxVADConfig,
    SherpaOnnxVADProvider,
)
from roomkit.voice.pipeline.turn import SmartTurnConfig, SmartTurnDetector
from roomkit.voice.stt import SherpaOnnxSTTConfig, SherpaOnnxSTTProvider
from roomkit.voice.tts import SherpaOnnxTTSConfig, SherpaOnnxTTSProvider

stt = SherpaOnnxSTTProvider(SherpaOnnxSTTConfig.from_local_model("zipformer-en-2023-06-26"))
tts = SherpaOnnxTTSProvider(
    SherpaOnnxTTSConfig.from_local_model("kokoro-int8-multi-lang-v1_0", voice="af_heart")
)
pipeline = AudioPipelineConfig(
    vad=SherpaOnnxVADProvider(SherpaOnnxVADConfig.from_local_model("ten-vad")),
    turn_detector=SmartTurnDetector(SmartTurnConfig.from_local_model("smart-turn-v3.2-cpu")),
)

The complete agent, with a local LLM through llama.cpp and RNNoise, is examples/voice_local_offline.py.

The models

Name Config Download What it is
zipformer-en-2023-06-26 SherpaOnnxSTTConfig 73 MB Streaming STT, English, zipformer transducer (int8)
zipformer-en-20M-2023-02-17 SherpaOnnxSTTConfig 44 MB Streaming STT, English, small (20M params); drops the first word of audio resampled from 48 kHz
zipformer-en-kroko-2025-08-06 SherpaOnnxSTTConfig 57 MB Streaming STT, English, Kroko
zipformer-fr-kroko-2025-08-06 SherpaOnnxSTTConfig 57 MB Streaming STT, French, Kroko (1.5 s tail padding)
nemotron-3.5-asr-streaming-0.6b-560ms-int8 SherpaOnnxSTTConfig 475 MB Streaming STT, multilingual, Nemotron 3.5; detects the language, stream_language="fr" fixes it
parakeet-tdt-0.6b-v3-int8 SherpaOnnxSTTConfig 487 MB Offline STT, 25 European languages, Parakeet TDT v3
whisper-tiny.en-int8 SherpaOnnxSTTConfig 104 MB Offline STT, English, Whisper tiny.en
whisper-base-int8 SherpaOnnxSTTConfig 161 MB Offline STT, multilingual, Whisper base; detects the language, language="fr" fixes it
moonshine-tiny-en-2026-02-27 SherpaOnnxSTTConfig 30 MB Offline STT, English, Moonshine v2 tiny; about 0.05 s per sentence on a CPU
moonshine-base-en-2026-02-27 SherpaOnnxSTTConfig 111 MB Offline STT, English, Moonshine v2 base
kokoro-int8-multi-lang-v1_0 SherpaOnnxTTSConfig 132 MB Kokoro v1.0, 54 named voices, 24 kHz
vits-piper-en_US-amy-low SherpaOnnxTTSConfig 67 MB Piper "amy", American English, 16 kHz
vits-piper-fr_FR-siwis-medium SherpaOnnxTTSConfig 67 MB Piper "siwis", French, 22 kHz
kitten-nano-en-v0_8-int8 SherpaOnnxTTSConfig 31 MB Kitten nano v0.8, English, 8 named voices, 24 kHz
kitten-mini-en-v0_8 SherpaOnnxTTSConfig 68 MB Kitten mini v0.8, English, 8 named voices, 24 kHz
ten-vad SherpaOnnxVADConfig < 1 MB TEN-VAD
silero-vad SherpaOnnxVADConfig < 1 MB Silero VAD
gtcrn-simple SherpaOnnxDenoiserConfig < 1 MB GTCRN speech enhancement
dpdfnet2 SherpaOnnxDenoiserConfig 11 MB DPDFNet2, streaming, for a 16 kHz pipeline
dpdfnet2-48khz-hr SherpaOnnxDenoiserConfig 11 MB DPDFNet2, streaming, for a 48 kHz pipeline
smart-turn-v3.2-cpu SmartTurnConfig 9 MB Smart Turn v3.2, CPU build (int8)
smart-turn-v3.2-gpu SmartTurnConfig 32 MB Smart Turn v3.2, fp32 build for provider="cuda"

The offline STT models (Parakeet, Whisper, Moonshine) transcribe each utterance once it ends, behind a VAD; the streaming ones write partials as the person speaks. Each entry was run once before it went in: the French Piper voice, for instance, is transcribed word for word by Parakeet and Kroko.

python -m roomkit.local_models list prints the same table and marks the models already downloaded; --kind sherpa-onnx-tts narrows it to one config. From Python, list_local_models() returns them as LocalModels (name, kind, description, size, files, voices), and is_downloaded() says which are there:

from roomkit.local_models import SHERPA_ONNX_TTS, is_downloaded, list_local_models

for model in list_local_models(SHERPA_ONNX_TTS):
    print(model.name, model.size_mb, is_downloaded(model), model.voices[:3])

Downloading

download fetches each file, checks it against its pinned SHA-256 and extracts the archives, in a staging directory renamed into place only when complete: an interrupted download leaves nothing half-written, and a file that does not match is refused. A model already there is skipped, so the command is safe to re-run. When RoomKit pins newer files for a model, the old copy reads as not downloaded and the next download replaces it.

Models live in $ROOMKIT_MODELS_DIR when it is set, else in ~/.cache/roomkit/models ($XDG_CACHE_HOME/roomkit/models). Every command and from_local_model() take --models-dir / models_dir= to use another one.

In a container, download at build time and the image runs offline:

ENV ROOMKIT_MODELS_DIR=/models
RUN python -m roomkit.local_models download zipformer-en-2023-06-26 kokoro-int8-multi-lang-v1_0 ten-vad

python -m roomkit.local_models path <name> prints a downloaded model's directory, for a script that needs the files themselves.

From Python, download_local_model(name) does the same as the command. It blocks while it downloads: call it at setup, not from an event loop.

Building a config by name

from_local_model(name, ...) exists on SherpaOnnxSTTConfig, SherpaOnnxTTSConfig, SherpaOnnxVADConfig, SherpaOnnxDenoiserConfig and SmartTurnConfig. It fills in the file paths and the settings the model needs (its mode, model type, sample rate), and any other field is passed as a keyword and wins over the model's settings:

SherpaOnnxSTTConfig.from_local_model("zipformer-en-2023-06-26", provider="cuda", num_threads=4)

It never touches the network. A model that is not downloaded raises ModelNotDownloadedError, whose message is the command that downloads it:

Local model 'ten-vad' is not downloaded in /home/me/.cache/roomkit/models.
Download it once, while online: python -m roomkit.local_models download ten-vad

A name that does not exist, or a model for another config (ten-vad given to SherpaOnnxSTTConfig), raises ValueError listing the names that fit.

Voices by name

A Kokoro voice is named rather than numbered. voice= sets the default voice, and a voice name also works per synthesis, a VoiceChannel(voice_map=...) entry included:

config = SherpaOnnxTTSConfig.from_local_model(
    "kokoro-int8-multi-lang-v1_0", voice="ff_siwis", lang="fr"
)
config.speaker_id                     # 30
await tts.synthesize("Bonjour", voice="ff_siwis")

The names come from the model's own speaker table and fill SherpaOnnxTTSConfig.voice_names; a speaker id given as a number still works. KOKORO_V1_VOICES in roomkit.local_models lists the 54 Kokoro names in id order, KITTEN_V08_VOICES the 8 Kitten ones (expr-voice-2-f...). A voice keeps its accent: pick one of the text's language, and set lang for a language other than English or Chinese.

RNNoise without a system library

RNNoiseDenoiserProvider loads librnnoise. When the system has none, it uses the copy the pyrnnoise wheel ships for Linux (x86_64, aarch64), macOS and Windows:

pip install roomkit[rnnoise]

Only the native library is loaded; the pyrnnoise package is never imported. Its declared dependencies (audiolab, matplotlib) are installed all the same. A system copy (apt install librnnoise0, or a source build) is preferred when present. Each pyrnnoise release ships a macOS wheel for the newest macOS only; on macOS 15, pip installs 0.4.3, the last release built for it.

What is not here

  • LLMs. LlamaCppAIProvider downloads its GGUF model itself, through llama-server, on first start; pass the path of a .gguf file to run one you already hold.
  • Models without an entry. Any sherpa-onnx or Smart Turn model still works through the config's path fields; the list holds the models RoomKit has checked end to end.