transcribe.cpp

Mozilla AI Releases Llamafile 0.10.4 With New Transcribefile Built On Transcribe.cpp

Mozilla AI vừa phát hành Llamafile 0.10.4 với tính năng mới Transcribefile, dựa trên thư viện Transcribe.cpp (tương tự Llama.cpp nhưng dành cho speech-to-text). Phiên bản này hỗ trợ 16+ mô hình TTS, cập nhật Llama.cpp, cải tiến tăng tốc Vulkan/AMD ROCm, bổ sung tải xuống HTTPS và sandboxing pledge/SECCOMP.

Là người phát triển AI hoặc chuyên về xử lý âm thanh và GPU, bạn nên đọc để khám phá cách Transcribe.cpp và Llamafile tối ưu hóa hiệu suất inference âm thanh với các mô hình lớn bằng cách tích hợp công nghệ Vulkan, CUDA và Apple Metal, giúp tiết kiệm thời gian và nguồn lực cho ứng dụng của mình.

Vocalinux 0.14 Beta Released For Offline Voice Dictation / Speech-To-Text On Linux

Vocalinux 0.14 beta vừa ra mắt, cung cấp tính năng nhận diện giọng nói offline (speech-to-text) trên Linux, hỗ trợ cả X11 và Wayland. Phiên bản mới bổ sung phím tắt cấu hình, hỗ trợ API từ xa FunASR/SenseVoice, cải tiến Wayland và tối ưu Whisper.cpp trên laptop lai, sử dụng Whisper.cpp (hoặc VOSK/OpenAI Whisper/PyTorch) với tăng tốc GPU Vulkan.

Nếu bạn là lập trình viên muốn phát triển ứng dụng xử lý giọng nói offline trên Linux, Vocalinux 0.14 beta sẽ cung cấp mã nguồn mở, backend linh hoạt (Whisper.cpp, VOSK, OpenAI) và hỗ trợ GPU/CPU hiệu quả, giúp bạn nhanh chóng tích hợp và tối ưu hóa giải pháp cho các dự án riêng của mình.

Introducing Real World VoiceEQ: Measuring the human quality of voice AI

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

AI-Enabled Advisory Services for Higher Education

Educational institutions with large call centers face a scalability problem: reviewing advisor quality across financial aid, admissions, and enrollment calls is manual, slow, and expensive. With Databricks, GenAI can ingest calls, transcribe them with high fidelity, score advisor performance against an institutional rubric, and surface insights through natural language, all on a single governed platform. This blog walks through the approach with code notebooks and an example Genie space.

Apple's New Speech API vs Whisper: The First Real Benchmark

A benchmark comparing Apple's new SpeechAnalyzer API (iOS/macOS 26) against the legacy SFSpeechRecognizer and three Whisper model sizes (Tiny, Base, Small) on 5,559 LibriSpeech utterances. SpeechAnalyzer achieved 2.12% WER on clean speech and 4.56% on noisy speech, beating Whisper Small (3.74%/7.95%) while running ~3x faster. The legacy SFSpeechRecognizer came last at 9.02%/16.25%. Results were validated against OpenAI's published Whisper numbers, and raw transcripts are publicly downloadable. The conclusion: developers should migrate from SFSpeechRecognizer immediately, and for English on Apple hardware, SpeechAnalyzer is now the strongest on-device option. Whisper retains advantages in language coverage (100+ vs ~30 locales) and cross-platform support.

Pluralistic: Reverse centaurs are the answer to the AI paradox (11 Sep 2025) – Pluralistic: Daily links from Cory Doctorow

Cory Doctorow introduces the concept of 'reverse centaurs' to explain why some people find AI empowering while others find it oppressive. A centaur is a human assisted by a machine on their own terms; a reverse centaur is a human puppeteered by a machine, serving as an 'accountability sink' for AI errors while being given impossible workloads. The Hearst summer reading guide debacle — where one writer was tasked with producing a 64-page AI-generated supplement — illustrates the reverse centaur dynamic. Doctorow argues the AI bubble will burst, leaving behind open source models like Whisper, cheaper GPUs, and better-trained engineers, but also massive economic damage and 'digital asbestos' embedded in society's infrastructure. He is writing a book titled 'The Reverse-Centaur's Guide to AI' aimed at improving AI criticism and deflating the investment bubble.

Silent speech with ultrasound — Aleph

A team trained a model to decode speech from ultrasound video of the tongue during silent articulation, achieving a 15.6% word error rate on open-vocabulary speech — approaching lip-reading benchmarks despite using only 50 hours of training data collected over one month. The system uses a ResNet-18 2+1d video encoder aligned to Whisper's embedding space, enabling the Whisper decoder to transcribe tongue movements into text. Notably, the model generalizes to new speakers without retraining. Data was collected using vocalized speech (to enable audio quality checks) with the assumption that tongue movements transfer to silent speech. The two main remaining hardware challenges are miniaturizing the ultrasound probe and replacing gel coupling with hydrogel for practical wearable use.