All Audio items under Apps on AIFIRST News, newest first. 17 items.
audio.cpp has released quantized versions of YuE2, an open-source music generation model, in its dev branch. Q4 and Q8 weight variants reduce VRAM requirements to 8-9 GB, making the model accessible on consumer GPUs. Prebuilt binaries are available via the Artifacts section of the repository.
Orukeet is a new automatic speech recognition model derived from Parakeet, surfaced as a recommendation within OpenWhispr. It is positioned as a more accurate alternative to Whisper variants for conversational English. Formal benchmarks are not yet published, but early user reports suggest lower word error rates.
YuE2 is a new music generation model that takes lyrics and a style prompt, generates a melody-and-chord plan, then produces a complete song with vocals and accompaniment. The model targets frontier song quality and is designed as a significant improvement over the original YuE release.
An independent researcher trained an audio model that generates one-shots for music production from text prompts. The model treats timbre and instrument response as separate controllable parameters rather than relying on high-level genre descriptions.
Audacity4-MCP is an early-alpha MCP server that connects AI assistants to Audacity 4 for AI-driven audio editing. The project allows MCP-compatible tools like Claude to control Audacity programmatically, enabling AI-assisted audio workflows.
A developer released sanoTTS, a text-to-speech system with 11 voices and 6 languages packed into 294k parameters (337 KB). The model runs on microcontrollers with 512 KB of SRAM and no NPU, such as chips costing around $3. This makes it one of the smallest functional multilingual TTS systems available.
Ali is an open-source AI singing voice model trained on the DiffSinger diffusion-based architecture and released as a public archive for research and creative use. DiffSinger synthesizes singing voices from music scores with high naturalness. The archive lets artists and researchers work with a trained voice model without training from scratch.
StemDeck is a free, open-source tool that separates audio tracks into individual stems (vocals, drums, bass, other) entirely on the user's local machine with no cloud subscription required. It runs the separation model offline, offering privacy and low latency for musicians and producers. The project ships a cross-platform desktop interface and is under active development.
audio.cpp version 0.7 has been released, expanding the project to 62 audio model families with over 85 model variants available for local inference. New in this release is an Arena UI that lets users compare models side by side locally without sending audio to cloud services, and support for recent MiniMax audio models. audio.cpp follows a similar architecture to llama.cpp but focuses entirely on audio generation, voice, and speech models.
A developer trained a 1.2B parameter diffusion transformer game music generator from scratch on a cloud H100 over 8 days, using the VAE from Stable Audio 3. The model, named Localsong, generates instrumental game music and is publicly available on Hugging Face.
Lyricist is a free AI songwriting studio for Windows distributed as a downloadable release, with no subscription required. The tool targets musicians and songwriters who want AI writing assistance without ongoing cost.
A public instance of Muscriptor, described as the most capable current Audio-to-MIDI conversion model, is now accessible at pianoify.net. The service converts audio recordings to MIDI notation without local installation.
Nari Labs describes five inference optimizations applied to Qwen3-TTS on a single NVIDIA H100 SXM, reaching 10 requests per second and sub-50 ms p95 time to first audio. Throughput is about 630 characters per second at full load, at roughly $2 per 1 million characters. The post benchmarks the result against vLLM-Omni, SGLang-Omni, and VoxServe, and against the commercial services ElevenLabs and Cartesia, claiming lower latency at higher request volumes.
A 125-million-parameter decoder-only transformer was trained to autocomplete piano melodies in real time, reaching 108 note predictions per second on an iPhone 15. The key design choice was a unified note token format encoding pitch, timing, duration, and velocity together, so the model advances by complete notes rather than individual attributes. Post-training with Direct Preference Optimization using Gemini-based pairwise evaluation boosted preference scores from 24.5% to 69%, and the model powers the RollTab iOS app.
QMind applies quantum mechanics math (superposition, interference, wavefunction collapse) to AI reasoning on a standard CPU - no quantum hardware. Instead of one reasoning path, it explores many simultaneously; paths ...
YuE2-3B, a new open music generation model, has been released with an available demo site. Community members report solid output quality for a 3 billion parameter model in the music generation space.
AIFIRST 
