Hugging Face's speech-to-speech is a self-hostable, OpenAI-Realtime-compatible voice agent pipeline that swaps out per-minute fees from platforms like Vapi and Bland AI.
Run an AI voice agent for your business today, an AI receptionist, an outbound sales dialer, a customer service line that never puts anyone on hold, and you are paying by the minute for the privilege. Vapi's pricing page lists calls at $0.05 per minute for hosting, plus $10 a month for every concurrent line, plus whatever your speech-to-text, language model, and text-to-speech providers charge on top, since those costs are passed through at cost rather than bundled in. Bland AI charges a flat $0.11 to $0.14 per minute depending on tier, plus a $299 or $499 monthly platform fee once you outgrow its free-to-start plan. Hugging Face's speech-to-speech, a project that landed on GitHub's trending list this week with 7,800 stars, is the same underlying pipeline, voice detection, transcription, a language model, and speech synthesis, packaged as software you run yourself under an Apache-2.0 license. No per-minute meter, no platform fee, just the cost of the compute.
The repo is a four-stage cascade: Silero VAD detects when someone starts and stops talking, a speech-to-text model transcribes the turn, a language model generates the reply, and a text-to-speech model speaks it back, with every stage running in its own thread and swappable via command-line flags. The part that actually threatens the Vapi and Bland business model is the API surface, not the model quality: the server exposes an OpenAI Realtime-compatible WebSocket endpoint, the same protocol shape that hosted voice platforms and OpenAI's own Realtime API use. A team that already built a voice agent against that protocol can point its client at a self-hosted server instead and keep most of its existing integration code.
It is also not locked into any one vendor for the expensive part. The language model slot speaks OpenAI-compatible protocols, so you can run it against OpenAI, against Hugging Face's own Inference Providers, against OpenRouter, or against a vLLM or llama.cpp server on hardware you already own, for a setup with no external API bill at all. The project's own documentation points to Reachy Mini, a real production robot, as a live user of this exact pipeline, which is a stronger production-readiness signal than most trending repos offer.
The audience for this is narrower than "every business," but real. A founder who has already built an outbound sales or appointment-reminder line on Vapi or Bland and wants to cut the per-minute bill once volume gets meaningful. An internal IT or ops team standing up an AI receptionist who has a person comfortable running a pip install and reading CLI flags. An agency running voice agents for several clients that wants to stop paying per-minute margins on every one of them. It is not a fit for someone who wants a dashboard and a phone number in five minutes with no infrastructure to think about.
Here is the honest part: self-hosting is not free, it just moves the cost from a line item to a headcount problem. You need a machine running continuously, ideally with a GPU, and someone accountable for keeping it patched and scaled as call volume grows. The pipeline itself does not include telephony, there is no built-in way to actually receive a phone call, so you still need to wire up Twilio or a SIP provider yourself, work that Vapi and Bland both bundle or make close to turnkey. There is also no HIPAA or SOC 2 certification bundled in, so if you need what Vapi sells for its $2,000-a-month HIPAA add-on, that compliance work becomes yours to do, not a checkbox. And with 82 open issues and 33 open pull requests at last check, this is an actively maintained project, not a finished product you install and forget.
The per-minute meter on hosted voice platforms was never really pricing the difficulty of the AI, it was pricing someone else's decision to run the infrastructure for you. Hugging Face just handed you the infrastructure.