NPort is a lightweight ngrok alternative that creates secure HTTP/HTTPS tunnels from your localhost to public URLs using Cloudflare. No configuration, no accounts, just instant tunnels with custom subdomains.

https://github.com/tuanngocptn/nport https://x.com/tom_doerr/status/2083068827177222388

build real-time voice agents using open-source models!

Hugging Face just open-sourced speech-to-speech, and it gives you the entire voice pipeline as four swappable models.

VAD → STT → LLM → TTS

↳ VAD (voice activity detection) watches the microphone stream and decides when you started and stopped speaking

↳ STT (speech to text) turns that segment of audio into a transcript

↳ the LLM reads the transcript and streams back a reply as text

↳ TTS (text to speech) turns that text into audio and streams it to your speakers

each stage runs in its own thread and passes work forward through a queue, so nothing sits waiting for a full turn to finish.

the alternative is one model that takes audio in and gives audio out. simpler to run, but you inherit its voice, its language coverage, and its latency.

here you choose each part. Parakeet or Whisper for transcription, Qwen3-TTS or Kokoro for the voice, and any OpenAI-compatible model for the reasoning, hosted or running locally through llama.cpp.

the whole pipeline is served as an OpenAI Realtime-compatible WebSocket API, so an existing Realtime client works by changing one URL.

now the hard part.

when you speak over the assistant, three things are already in motion. LLM tokens are streaming, TTS audio is playing, and more chunks are sitting in the send queue.

all of it has to disappear without leaving the pipeline broken.

their fix is one shared counter. every response carries the counter's current value, and the LLM and TTS threads check it against the counter on every token they produce.

cancelling increments the counter, which makes every response in flight stale on its next check. a separate flag drops the audio that was already queued.

no timers, no race between two separate signals.

this isn't a demo either. the same pipeline runs in production as the conversation backend for thousands of Reachy Mini robots.

link to the repo: https://github.com/huggingface/speech-to-speech

i have previously recorded a video on building a real-time voice agent. the same is quoted below.

stay tuned, i'll be recording a new one using the HF speech-to-speech stack.

cheers! :)

Tailscale's open-source core gives you the daemon and CLI to build private WireGuard networks without the complexity, running on Linux, Windows, macOS, FreeBSD, and OpenBSD. It's the same code powering the mobile apps, minus the GUI.

Explore it here:
https://www.opensourceprojects.dev/post/tailscale https://x.com/GithubProjects/status/2083033555475812505

seedance 2.5 me tiene loco

físicas que se sostienen, cámara que se mueve como si la manejara alguien real, personajes que no pierden consistencia a los 30 segundos

esto ya no parece IA

seedance 2.5 llegara pronto a higgsfield https://x.com/robiartec/status/2083199742469636275