Kyutai Moshi vs so-vits-svc

A side-by-side look at pricing, capabilities, pros, cons, and our editorial scores.

	Kyutai Moshi Audio	so-vits-svc Audio
Tagline	Open-source, full-duplex speech-to-speech foundation model with sub-200ms latency	SoftVC VITS Singing Voice Conversion — open-source pipeline for training and running singing-voice models.
Category	Audio	Audio
Pricing	Free· Free and open source. Models under CC-BY 4.0, code under MIT (Python) / Apache 2.0 (Rust). Self-hosted only — you pay your own compute (24GB+ GPU for PyTorch, or Apple Silicon via MLX).	Free· Free / open-source (AGPL-3.0). You provide your own compute (typically a CUDA-capable GPU) and training datasets.
Model	Moshi (7B-class speech-text foundation model) + Mimi neural audio codec, in-house by Kyutai	SoftVC content encoder + VITS backbone + NSF-HiFiGAN vocoder; optional ContentVec, HuBERT-Soft, Whisper-PPG, WavLM encoders and shallow-diffusion module.
Editorial score	—	—
Use cases	Real-time voice assistant prototypesResearch on full-duplex spoken dialogueOn-device voice interaction on Apple Silicon via MLXLow-latency conversational agents behind WebSocketNeural audio codec experimentation with MimiSelf-hosted voice interface for privacy-sensitive appsSpeech tokenizer for downstream audio LLM trainingInterruptible in-car or wearable voice UX	Singing voice conversion (AI covers)VTuber and virtual-character singing voicesCustom vocal timbre for indie music productionSpeaker mixing and timbre morphing experimentsVoice model training on curated datasetsResearch on VITS-based voice synthesisONNX export for lightweight SVC inference
Pros	Truly full-duplex — handles interruptions, overlap and back-channels rather than rigid turn-taking Sub-200ms practical latency on a single L4 GPU, well below third-party voice APIs Fully open weights (CC-BY 4.0) plus MIT/Apache code — self-host with no per-minute billing Ships with Mimi, a streaming neural audio codec that beats SpeechTokenizer and SemantiCodec Multiple inference backends: PyTorch for research, Rust/Candle for production, MLX for on-device Mac/iPhone Inner-monologue text prediction gives you a transcript alongside the audio stream for free	Fully open source (AGPL-3.0) and runs entirely offline — no per-use fees, no data leaving your machine. State-of-the-art singing quality for its generation: NSF-HiFiGAN vocoder + shallow diffusion noticeably reduce breath and sibilance artifacts. Pluggable content encoders (ContentVec, HuBERT-Soft, Whisper-PPG, WavLM) let you trade off timbre leakage vs. pronunciation fidelity. Speaker mixing (static and dynamic) and clustering-based timbre control give producers real creative knobs beyond one-shot conversion. ONNX export enables inference on non-PyTorch runtimes and lighter deployment targets. Huge community: 28k+ GitHub stars, dozens of active forks, tutorials, and ready-made WebUI front-ends.
Cons	English-only voices at launch — no multilingual support out of the box Knowledge and reasoning quality trail top text LLMs; it's a 7B-class model, not GPT-4o Voice Requires a 24GB+ GPU for the reference PyTorch build; on-device is only viable via MLX on Apple Silicon No hosted API or SaaS tier — you own the ops, scaling and safety filtering Only two fixed synthetic voices (Moshiko/Moshika); no voice cloning or speaker conditioning in the release	Upstream repo is archived (Nov 2023) — no official updates, security fixes or new-model support; you rely on forks. Steep setup: CUDA GPU, correct PyTorch/torchaudio versions, manual f0 extraction and slicing, and hours-to-days of training per voice. No built-in UI or hosted inference — you either script it or bolt on a third-party WebUI. Documentation is uneven and partly Chinese-first; several config knobs (diffusion depth, cluster ratio, encoder choice) require trial and error. Serious ethical / legal exposure: cloning a real person's singing voice without consent runs into copyright, publicity-rights and (increasingly) deepfake-specific laws — the license does not absolve you.
Website	kyutai.org	github.com

Pick Kyutai Moshi if

✅ Truly full-duplex — handles interruptions, overlap and back-channels rather than rigid turn-taking
✅ Sub-200ms practical latency on a single L4 GPU, well below third-party voice APIs
✅ Fully open weights (CC-BY 4.0) plus MIT/Apache code — self-host with no per-minute billing
✅ Ships with Mimi, a streaming neural audio codec that beats SpeechTokenizer and SemantiCodec

Pick so-vits-svc if

✅ Fully open source (AGPL-3.0) and runs entirely offline — no per-use fees, no data leaving your machine.
✅ State-of-the-art singing quality for its generation: NSF-HiFiGAN vocoder + shallow diffusion noticeably reduce breath and sibilance artifacts.
✅ Pluggable content encoders (ContentVec, HuBERT-Soft, Whisper-PPG, WavLM) let you trade off timbre leakage vs. pronunciation fidelity.
✅ Speaker mixing (static and dynamic) and clustering-based timbre control give producers real creative knobs beyond one-shot conversion.

Compare a different pair →