AI News

Weights AI is a browser-based AI cover generator that changes the singing voice on an existing song while keeping its melody, lyrics, and timing. Users can choose from more than 2,000 community voice models, upload a song, and receive a cover in about a minute. Supported uploads include MP3, WAV, M4A, AAC, FLAC, and OGG files up to 50 MB, and users can also record vocals in the browser. The service separates vocals from the instrumental track, re-sings the vocal with the selected voice, and mixes the result back with the music; a pitch-matching setting is available when the new voice has a different range. The voices are RVC models trained on recordings of singers, characters, or speakers, and real-person voices are described as unofficial fan models. Users may instead train a private model from at least 10 minutes of a cappella singing; training takes about 15 to 30 minutes. Signing in provides 10 free credits, enough for one cover, while each cover uses 10 credits. Finished covers can be downloaded as MP3 and saved in the user’s history. Weights AI says real-person voice covers are for personal, non-commercial use, while paid-credit remixes made with an original AI voice can be used commercially.

TurboScribe AI is a free web tool that converts audio and video recordings into editable, timestamped transcripts and subtitle files. It downloads a compact Whisper-based speech model to the browser, then performs transcription locally through WebAssembly on the device's CPU or GPU; the recording itself is not uploaded. The first use downloads and caches a model of about 39 MB, after which files are processed in 30-second pieces and the text appears progressively. Video formats including MP4, MOV, WEBM, MKV, and AVI have their audio tracks extracted automatically, while supported audio includes MP3, WAV, M4A, FLAC, OGG, AAC, and OPUS. Users can select a spoken language or use auto-detection, edit the transcript, and copy or export it as TXT, SRT, or VTT. The service requires no account, email address, credit card, or per-minute payment, but processing speed and practical file length depend on the user's hardware because there is no server-side processing limit. Long recordings are divided into ten-minute reading segments and may run more reliably on a desktop or laptop with more memory. Local processing stops if the browser tab is closed or reloaded. The compact browser model is intended to be fast and private, but it is less capable than larger server-side models on difficult audio. Background noise, room echo, overlapping speakers, strong accents, names, and numbers can reduce accuracy, so the article recommends setting the language when possible and reviewing the transcript before export. The tool lists 15 language options, with English described as the most dependable, and is positioned for meetings, interviews, lectures, podcasts, voice memos, and caption creation.

Granite Switch 4.1 is presented as a single endpoint for chat, decision-making, and speech. The page also lists Granite Speech 5.0-470m-turboctc, an automatic speech recognition model with 0.5B parameters. The listed model was updated 14 days ago, and the page displays additional figures of 59.1k and 80 without identifying their labels in the extracted text.

Sunny is an open-source Electron desktop assistant designed for older adults, non-technical users, and people who want to automate computer tasks without writing scripts. Users can type or speak requests, while Sunny uses screen understanding to identify windows, buttons, text fields, and navigation elements, then move the mouse, click, type, scroll, and open applications with permission. Its accessibility features include high-contrast visuals, enlarged text, text-to-speech, voice input, hands-free cooking instructions, medication reminders, and spoken weather and news digests. The project offers verified recipe search and a cooking mode with large step-by-step controls, scheduled vocal medication alerts, and a local weather and headline digest. AI can run offline through an embedded Moondream vision model or through Ollama with local vision-capable models; OpenAI, Anthropic, and Groq are optional cloud providers that require explicit consent. Sunny says screen captures, preferences, API keys, and logs remain on the device by default, with API keys protected by the operating system’s secure storage. Destructive actions such as sending email, deleting files, making purchases, or changing settings require confirmation. The assistant also sanitizes screen text and detects possible indirect prompt injections, stopping when it identifies a high-severity threat. Pre-built installers are provided for Windows, macOS, and Linux, and the project is released under the MIT License.

Veda is a voice agent designed to remain present on the desktop as a small taskbar ring, waking when users say its name or press Win+Space. It can converse, answer follow-up questions, and operate applications, files, windows, media, settings, timers and reminders through keyboard- and mouse-equivalent actions. Speech recognition, conversation handling and voice are supplied by Deepgram services; the default configuration uses Nova-3 for recognition, Flux for turn detection, Aura-2 Helena for speech, and OpenAI’s gpt-4.1-mini as the language model, while settings can be changed through the Agent panel. The agent keeps a microphone listener open for wake-word detection, but local voice activity detection prevents audio from being sent while nobody is speaking. Recognition includes pre-roll audio, background-noise handling, wake-name matching and a ten-minute-per-hour cap. Once awake, a Voice Agent conversation streams 16 kHz microphone audio and 24 kHz voice output, ducks other audio, buffers network bursts and flushes queued speech when the user interrupts. Echo cancellation and a separate echo gate reduce the risk that the agent hears its own voice, while still allowing louder overlapping speech through. Veda exposes tools for desktop status, applications, windows, files, volume, Wi-Fi, wallpaper, notifications, timers, memory and system tasks. Routine actions run immediately; sensitive and destructive actions require approval, with destructive actions confirmed every time and requests expiring after three minutes. Process identity limits which components can approve requests or change settings, and application content is treated as data rather than instructions. Personal memory and recent conversation information remain in a private local JSON file, while the API key is kept in a private directory and excluded from the model prompt. Applications can become compatible by registering descriptions, state and explicit actions through the platform’s agent interface. The project includes host tests and deterministic agent simulators, but real Deepgram tests require an API key and incur charges. Current limitations include English-only operation, possible speech-recognition errors, dependence on Internet access and a Deepgram account, and reduced ability to interrupt the agent when speakers are very close to the microphone.

Apple announced the fourth-generation Apple TV 4K, adding a full-screen Siri AI experience powered by the next generation of Apple Intelligence. The assistant is designed to search for and recommend entertainment conversationally, answer follow-up questions about shows, use personal context from an Apple Intelligence-enabled iPhone, and take more systemwide app actions. Three far-field microphones and echo cancellation enable hands-free “Siri” or “Hey Siri” recognition while audio is playing. Apple says the A19 chip provides up to 50 percent faster CPU performance and up to twice the GPU performance of the previous generation, and adds a dedicated AV1 decoder for more efficient streaming. The new Siri Remote adds Find My support, Precision Finding, and a sound option for locating the remote. The device can create a four-speaker surround setup with four HomePod mini 2 speakers or matching front and rear pairs of HomePod (2nd generation), with automatic room-based audio calibration. As a home hub, it supports smart-home controls, automations, doorbell interactions, Wi-Fi, Thread, and Apple Intelligence-generated camera activity descriptions and summarized notifications. Siri requests can use on-device processing or Private Cloud Compute; Apple says data processed there is not stored or accessible to Apple. The company also describes Secure Exclaves for isolating microphone data and end-to-end encryption for HomeKit Secure Video. Apple TV 4K can be pre-ordered for $299 in the United States, with store availability beginning November 13. Siri AI is also offered as a beta on Apple TV 4K models from the second generation onward, initially in English across the United States, Australia, Canada, India, New Zealand, Singapore, South Africa, and the United Kingdom. Availability requires supported software and, for some personal-context and Home features, an Apple Intelligence-enabled iPhone or iPad; some server-side features have daily usage limits.

TurboScribe AI is a browser-based transcription studio that converts audio and video into editable, timestamped text. It says processing runs locally through WebAssembly using a Whisper-class neural speech model, so files are not sent to a server. The service accepts common audio formats such as MP3, WAV, M4A, FLAC, and OGG, as well as video formats including MP4, MOV, WEBM, and MKV, extracting audio from video automatically. Users can choose a spoken language or use auto-detection, with English, Chinese, Spanish, French, German, Portuguese, Italian, Dutch, Polish, Russian, Turkish, Japanese, Korean, Arabic, Hindi, and Indonesian listed as supported options. Transcripts can be edited and copied or exported as plain TXT, SRT subtitles, or VTT web captions. Processing is performed in 30-second chunks with progress updates, and the page says there is no signup, account, credit card, watermark, per-minute metering, or hard server-side file limit. The product is presented for meetings, interviews, podcasts, lectures, voice notes, and creator captioning. Accuracy is described as strongest for clear audio, while noise, overlapping speakers, and strong accents can reduce quality; longer files depend on the device's memory and processing speed.

ElevenLabs has launched Eleven v4, a text-to-speech model designed to interpret tone, pacing, emotion, character, and conversational context, alongside Eleven v4 Turbo for lower-latency applications. The company says Eleven v4 ranked first on Artificial Analysis’s Provider Voice Arena leaderboard and was preferred by about 75% of listeners in blind comparisons with competing systems, though these are company-reported evaluations. Eleven v4 Turbo has a median inference latency of about 100 milliseconds and a median time to first speech of about 150 milliseconds, according to measurements described in the article. Both models support natural-language delivery instructions, inline audio tags, and improved International Phonetic Alphabet pronunciation control. They are designed to produce more consistent multi-speaker dialogue and preserve speaker identity across conversations, narration, regenerated lines, and longer projects. Both models support more than 90 languages, including cross-language speech that retains the original speaker’s identity while adopting a native accent. Voice cloning can use 10 seconds of audio for Instant Voice Clones, while Professional Voice Clones are also supported. Eleven v4 and Eleven v4 Turbo are available through ElevenAgents, ElevenCreative, and the ElevenAPI.

NVIDIA’s Hugging Face page presents Magpie TTS Multilingual, a 0.2B-parameter text-to-speech model for low-latency multilingual voice agents. Its title highlights open weights and full deployment control. The listing was updated 23 days before extraction and showed about 11.3K uses and a displayed value of 199. The excerpt does not include benchmark results, supported languages, licensing terms, or deployment instructions.

Hugging Face’s Open TTS Leaderboard is presented as a scalable evaluation resource for multilingual text-to-speech and voice cloning. The available excerpt lists FunAudioLLM/Fun-CosyVoice3-0.5B-2512 as a text-to-speech entry. The entry was updated on February 3 and shows the figures 197k and 651, though the excerpt does not define those metrics.

ASTRO is an open-core, self-hosted voice AI agent designed to hear, speak, see, execute commands, and train LoRA adapters on its own device. The project targets an offline-first setup centered on a Raspberry Pi 5 with a Hailo-10H NPU, while PC hardware and cloud keys are optional. Its voice stack combines a wake word, Polish speech recognition through Vosk or Whisper, and Piper text-to-speech; its vision stack supports network cameras, object and face detection, person recognition, OCR, age and gender estimation, motion events, and scene description through a vision-language model. A local learning loop collects knowledge, distills trajectories, and trains adapters. The repository is organized into CPU, Hailo NPU, PC, and optional remote backends, with modules for intent dispatch, tools, audio, vision, memory, safety, and deterministic skills. SQLite stores learned knowledge, trajectories, sessions, and biometric data, with AES-GCM protection for biometric information. Safety rules keep commands out of the cloud, enforce refusals, and protect secrets. The code can fall back to CPU execution and simulated tests without the full hardware set, although the documented hardware provides the intended experience. The quickstart installs dependencies, creates a local machine configuration, runs hermetic tests, and launches a text REPL. The project describes roughly 990 standard-library tests, quality gates for model promotion, an MIT license, and a sanitized public mirror whose private core may be ahead of the snapshot.

The author describes building a spelling game that needed one clear WAV recording for every word, then working through the problems created by AI-generated speech. Recordings made with audio.cpp were produced in English, French, and Norwegian; the workflow moved from OmniVoice to VoxCPM2 for voice creation because of licensing requirements, and used a VoxCPM2 reference recording with Chatterbox to generate sentences. A carefully written teacher prompt and a reference sentence were used to encourage calm, deliberate articulation, with the reference later slowed to 85% of its original tempo. Generating isolated words proved unreliable: some outputs lost the beginning, ended too early, or included part of the preceding word. Adding a sentence template gave the model more context, but extracting the repeated final word remained difficult. Silence detection could identify quiet intervals, but could not determine whether a pause separated words or occurred inside a word. Rules based on the last or longest pause, remaining duration, letter count, or volume thresholds each failed on different examples. The author therefore built audio-lab, a browser tool for waveform inspection, zooming, slowed playback, manual selection, silence overlays, and loading machine-generated selection JSON. Estimating where the introductory sentence should end and searching nearby pauses improved the heuristics, but the author emphasizes that this was rule refinement rather than model training. Changing the sentence wording and punctuation also improved generation; in English, a template beginning with “Please say the word” worked best in the reported tests. A separate cleanup step removed some unwanted lead-in sounds, while preview WAV files and JSON made the result inspectable. Whisper and nb-whisper supplied useful mismatch signals for batch review, although transcription flags still required listening: seven Norwegian flags produced only one judged error, while seven English flags were accepted. Licensing the model weights remained important, with OmniVoice Word-Control marked CC-BY-NC-4.0 and the cited Chatterbox and VoxCPM2 model cards listing MIT and Apache-2.0 licenses. The article concludes that AI accelerated the creation of small diagnostic tools, but human listening and judgment remained necessary.