AuK

Audar-ASR-V1 is a family of Arabic-first generative speech-recognition models developed by AudarAI. It treats transcription as audio-conditioned next-token prediction with a language-model decoder, rather than a CTC or transducer objective, using a Whisper-style 128-mel audio encoder and a Qwen3 decoder with a 30-second context. The models cover Modern Standard Arabic, major Arabic dialects, code-switched Arabic-English, English, and 30 languages overall. The Flash tier is intended for real-time, edge, on-device, or offline use, while Turbo targets lower error on difficult dialectal and long-form audio. Both tiers share an architecture and prompt interface and can be run through Transformers, GGUF with llama.cpp, or vLLM; the repository provides model pointers, benchmarks, and inference examples. The published models use the AudarAI Open v1.0 and AudarAI Community v1.0 licenses.

Voicebox is a local-first, open-source AI voice studio developed in the jamiepine/voicebox repository. It combines voice cloning, text-to-speech, speech-to-text dictation, voice profiles, audio effects, and agent voice output in a desktop application. It generates speech through seven switchable TTS engines, including Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox, HumeAI TADA, and Kokoro, and supports voice cloning from reference audio, preset voices, multilingual generation, automatic text chunking, crossfading, and multi-track story editing. Whisper-based transcription powers global dictation, in-app recording, and the transcription API; an optional local Qwen3 LLM can refine dictation or rewrite text according to a voice profile's persona. The application exposes REST endpoints for speech generation, agent speech, transcription, and profile management, plus a built-in MCP server with tools for speaking, transcribing, and browsing captures and profiles. It runs models and stores voice data locally, using Tauri with a React/TypeScript frontend and FastAPI backend; platform backends include MLX on Apple Silicon and PyTorch on CUDA, ROCm, XPU, DirectML, or CPU. The repository is licensed under the MIT License and provides macOS, Windows, and Docker distribution, with Linux build-from-source instructions.

VoiceStudio is an open-source, fully local desktop application for voice cloning and design, text-to-speech, transcription, dictation, video dubbing, and long-form audio production. It runs on macOS Apple Silicon, Windows, Linux, or Docker and keeps voices, projects, settings, and generated files on the machine by default. The application combines a Tauri desktop shell, a React and Vite interface, and a local FastAPI backend with registries for multiple TTS and ASR engines. Its workflows can transcribe and translate video, preserve speaker assignments, synthesize replacement speech, render audiobook chapters, separate vocals, diarize speakers, process batch jobs, and route work to CPU, CUDA, Apple Silicon MPS/MLX, ROCm, or optional remote workers. It also exposes local REST, SSE, WebSocket, OpenAI-compatible audio, and MCP interfaces, including speech synthesis and transcription endpoints. VoiceStudio is licensed under AGPL-3.0; downloaded models and tokenizers retain their own licenses. The local workflow requires no account, API key, subscription, or usage meter, although remote workers, configured external ASR endpoints, analytics, and other network-backed functions are opt-in. The repository describes the software as an active beta and distributes packaged desktop releases as well as source code.

YuE2 is an open music-generation model and Python pipeline that turns lyrics and a style prompt into a symbolic melody-and-chord plan and then renders that plan as a complete stereo song with vocals and accompaniment. It supports zero-shot covers by using a transcribed melody score, and composition editing by allowing the score, lyrics, style, harmony, melody, tempo, or form to be revised before rendering. Its AR–NAR Mixture-of-Transformers backbone predicts symbolic scores and semantic tokens, generates acoustic latents with flow matching, and uses a VAE to decode them into 48 kHz audio. The staged API exposes plan(), generate_semantic(), synthesize(), and decode() operations; the repository also provides an agent skill for song generation, transcription, cover creation, ABC-score editing, and listening comparisons. The repository provides the YuE2-3B model and requires Linux, Python 3.12, and an NVIDIA GPU with BF16 support and 24 GB of VRAM for the documented quick start. First-party code, documentation, and the agent skill use Apache 2.0, while the model weights use CC BY-NC 4.0. The original YuE implementation is preserved on the repository's YuE-v1 branch.

VoxCPM is an open-source tokenizer-free text-to-speech system from OpenBMB for multilingual speech generation, voice design, and voice cloning. Its latest major release, VoxCPM2, directly generates continuous speech representations in the latent space of AudioVAE V2 through a four-stage diffusion-autoregressive pipeline: LocEnc, TSLM, RALM, and LocDiT. It accepts text, optional natural-language voice descriptions, or reference audio with transcripts, and produces streaming or batch speech audio; VoxCPM2 supports 30 languages, voice-style controls, controllable and continuation-based cloning, and native 48 kHz output. The project provides a Python package and command-line interface, a local web demo, fine-tuning through full supervised training or LoRA, and serving options including an OpenAI-compatible vLLM-Omni endpoint. Its code and model weights are released under the Apache-2.0 license. The project warns that generated voices can be misused for impersonation, fraud, or disinformation, and notes that voice-design and controllable-cloning results can vary between runs.