Diarized transcription, zero-shot voice cloning, long-form speech synthesis — running on Apple Silicon, Android, Windows, and embedded Linux. No cloud APIs, no per-minute pricing, no data leaving the device.
brew install speechimplementation("audio.soniqo:speech:0.0.9")
A real session running NVIDIA Nemotron VoiceChat 11B locally: a spoken turn, an interruption mid-sentence, and a complete MCP tool call into Apple Reminders. RTF 0.92 on an M5 Pro, nothing leaving the machine.
Say it — Android does it: a complete command loop from speech to a real device action and a spoken reply, on one phone in 1.2 GB of RAM. Silero VAD, Parakeet STT, FunctionGemma tool calls, and Pocket TTS.
A four-minute open-source library tour: realtime transcription with Nemotron Streaming, local speech-to-speech with PersonaPlex, and 48 kHz voice cloning with VoxCPM2 — every demo runs on the laptop.
VAD → STT → LLM → TTS on phone and Mac · ~1.2 GB iPhone · ~1.5 GB S23 · <4 GB desktop
The same VAD → STT → LLM → TTS voice-agent pipeline on iPhone, Galaxy S23, and a Mac — with measured on-device memory for each budget.
Each group spans several sub-use-cases stitched from Soniqo components. Drop in your audio, get conversation, transcripts, or generated speech back — locally, in real time.
Build voice-first interfaces — from full-duplex speech-to-speech to wake-word-driven compositional pipelines, all running locally.
Turn audio into structured text — realtime streaming for live captions and dictation, batch high-accuracy for archives, diarized to name each speaker.
Synthesize speech in any voice — clone a voice in seconds, narrate audiobooks for hours, or cast multi-speaker podcasts, fully offline.
The use-case pipelines above are stitched from these models. Pick a component to read its architecture, CLI, Swift API, and benchmarks. All run on Apple Silicon, most also on Android and Linux.
52 langs, RTF 0.06, 4-/8-bit
Native CoreML: 1.40% WER, 6 s model load
MLX INT5: 128K offline context, RTF 0.028
Small, medium, large-v3, and turbo exports
32× real-time on Neural Engine
120M, 25 langs, streaming partials + EOU, ~232 MB on-device
1,672 languages, 300M–7B
14 langs, INT5 default, RTF 0.015
8 langs, INT5 default, RTF 0.074
Word-level timestamps, 80 ms
Streaming with punctuation
9 langs, zero-shot cloning, bf16 / 8-bit
12 Hz codec LM, faster than real-time
48 kHz, 30 langs, voice design + cloning
Zero-shot cloning, emotion + tempo controls, RTF 1.0
99M, 31 langs, 44.1 kHz
CoreML voice cloning, TTFT 0.27s, RTF 0.59
600+ langs, NAR diffusion cloning
Hindi emotion TTS + raw reference cloning
Style markers + zero-shot cloning
Zero-shot cloning in 16 flow steps, RTF 0.57
4B conversational cloning, emotion tags, RTF 0.78
54 voices, iOS-ready, 10 langs
English, ~130 ms streaming TTFA
90-min podcasts / audiobooks
9 langs, 5 baked voices, streaming
IndexTTS2, CosyVoice, Chatterbox Flash, Qwen3-TTS ICL
SpeechBrain ECAPA, 107 languages, 21M
Masked WeSpeaker + native PLDA/VBx, ~32 MB
Community-1 + Pyannote + Sortformer
WeSpeaker / CAM++ for ID
Silero v6.2.1, Pyannote, FireRedVAD
KWS Zipformer, 26× real-time
44.1 kHz stereo, variable-length music
Text → 30 s music, RTF 0.36
Open-Unmix + HTDemucs v4, 4 stems
DeepFilterNet3, 48 kHz real-time
Two-stream echo cancellation, 16 ms latency
Sidon denoise + dereverb → 48 kHz
48 kHz audio super-resolution
Streaming on-device LLM
Structured tool / function-call grammar
Many-to-many translation, 400+ languages
Streaming speech translation, FR/ES/PT/DE → EN
Moshi-family full-duplex speech-to-speech
Asymmetric duplex speech-to-speech, INT5 / INT8