Open-source multi-modal speech LLM. Unified understanding + generation in one model - ASR, TTS, voice conversion, speech dialogue. Strong expressive control and paralinguistic features. Available in Mini (8B) and Full variants.
Step-Audio 2 Mini is a local speech model from StepFun. It is best suited for cloning, dialogue, emotion, streaming workflows. Check the license before commercial use.
gpuapple
pytorchsafetensors
Multi-speaker + voice conversion
low
Apache 2.0
2025-08
pip install step-audiocloningdialogueemotionstreamingmultilingual
NVIDIA's SOTA lightweight ASR - 0.6B params, #1 on Open ASR Leaderboard for English. TDT (Token-and-Duration Transducer) decoding makes it 50× faster than Whisper Large v3 on GPU. Real-time streaming with word-level timestamps.
Ultra-lightweight yet stunning quality. 82M params only - runs on CPU in real-time. Best quality-to-size ratio of any TTS model.
Production-grade streaming ASR from Kyutai (makers of Moshi). Delay-streaming transformer with 500ms latency, word-level timestamps, speaker diarization. Top of Open ASR Leaderboard for real-time French + English.
Iterative upgrade over the original F5-TTS. Faster convergence via improved flow-matching schedule, better Chinese prosody, cross-lingual cloning. Now with streaming inference and improved CFM sampler.
OpenAI's optimized Whisper v3 with 4 decoder layers instead of 32. 8× faster than Whisper Large v3 with only minor accuracy trade-off. 99 languages supported. New gold standard for fast local transcription.
Desktop app & orchestrator for local TTS - not a model. Provides a UI studio, voice profile management, and a local API. Generates audio via swappable backends (Qwen3 TTS, Kokoro, Piper, XTTS…). Think of it as a front-end shell that runs on top of your installed TTS models.