Small specialized models are bringing TTS, wake word, voice separation, retrieval clarification and embedding search to CPUs and consumer hardware, reducing the need for GPU-heavy datacenters.
The subscription economics of large LLMs is under pressure, while specialized models with a few billion parameters solve tasks like OCR and segmentation. The thesis: the future is local, with smaller models, easier training, and no data center power ...
IBM Research releases ClariEdge, a dataset of about 3,900 examples for spotting ambiguous queries in RAG. A lightweight LoRA obtained through fine-tuning reaches 96.6% accuracy and outperforms larger baselines. The signal: specialized local component...
IBM Research has released ClariEdge, a dataset of about 3,900 examples for recognizing ambiguous queries in RAG systems and generating clarification requests. A lightweight LoRA reaches 96.6% classification accuracy and beats prompted baselines from ...
Kitten-TTS 2 brings 1.58-bit ternary quantization to a 1.7 billion parameter model that runs in real time on CPU. Self-hosting becomes realistic without GPUs, shifting TCO from VRAM to software execution quality. For organizations under data sovereig...
A 305-million-parameter speech encoder, frozen or augmented with low-rank adapters, takes on Saudi and Gulf Arabic speech recognition. The frozen system is statistically comparable to Cohere Transcribe Arabic on CER and 3.4 times faster on CPU; the a...
KittenML has released code and weights for Kitten-TTS 2, a TTS model with 1.58-bit ternary quantization and 1.7 billion parameters. It runs in real time on CPU, performs zero-shot voice cloning from under five seconds of audio, supports more than 20 ...
An Apache-2.0 open source project offers an on-device voice front-end for local assistants: five models for text-based wake word, voiceprint, voice separation and personal VAD. The Rust engine runs in the browser, Python and Android; models take 7.5 ...
A local test with EmbeddingGemma2 shows semantic image search on a consumer GPU, indexing 3,003 screenshots in 510 seconds with a 330 MB Q8 model and 768-dimensional vectors. But retrieval quality depends heavily on query wording and corpus character...
The open source project Openspline adds an avatar layer to existing voice AI pipelines: from an image, it generates a talking face with synchronized lips, executed locally on a GPU. The lite model exceeds 90 FPS on an NVIDIA RTX 4090 and integrates w...
