What we learned building a phone line that interprets calls in 47 languages

Hi folks, Alex here from ThunderPhone. We made a phone number that interprets conversations in both directions. You can merge it into a phone call or put it on speakerphone for an in-person conversation. It's (228) 888-8666, and there's more info at translatemycall.com. No app or account needed, and it's free for now. We first built this as a hackathon project in 2024. That version worked, but it was pretty rough. We recently realized that the stack we've built since then for ThunderPhone, our voice AI platform for production phone agents, was mature enough to rebuild it properly. The main problems ended up being: 1) Turn-taking is usually harder than the actual translation. A pause doesn't always mean someone is done talking, and people pause differently depending on the language. We run VAD separately for each speaker and tune the end-of-turn detection by language so it doesn't constantly cut people off or wait forever. 2) Phone audio loses a ton of information. Calls are generally 8 kHz narrowband audio, with compression, background noise, and sometimes two people talking at once. We run multiple transcription models at the same time and fuse their output with audio piped directly into LLMs. Either path can get something wrong, so using both helps quite a bit. 3) Every step has to stream. The call comes in over SIP and goes through a streaming pipeline for transcription, translation, and speech. There's still a beat of latency before it speaks, and heavy cross-talk can throw it off, but it feels much closer to a normal interpreted call than the 2024 version did. 47 languages are fully supported. Around 70 usually work overall, although quality drops off for some of the long-tail languages. If anyone wants to try it, the number is (228) 888-8666 and the site is translatemycall.com. Happy to answer questions about the pipeline or ThunderPhone.