No comments yet.
I recently finished what started as a hackathon-style project with a friend, which has now become TranslateMyCall.com.
The problem we set out to solve is that there’s currently no way to interpret phone calls between languages without a human in the loop. Apps like Google Translate work well for in-person conversations, and Skype (surprising, right?) is the best I’ve found for VoIP interpreting. But for traditional phone calls, your only option is paying $1 to $4 per minute for a human interpreter through a service like LanguageLine. Even then, it can take 5-10 minutes to get connected, and interpreter availability/quality can vary.
We thought building an AI app to interpret phone calls between languages would be simple (famous last words...) but quickly learned why nobody else has done it. Building telephony integrations through Twilio was a huge pain, and phone audio is low-quality (8KHz/8-bit), making high-accuracy transcription difficult, especially for non-English languages.
In fact, transcription turned out to be the crux of this whole project. Translating transcribed text and synthesizing translated speech were easier, but balancing transcription quality with low latency became our #1 challenge. High transcription accuracy is possible if you’re willing to wait 10+ seconds with models like Google’s Chirp 2, but for real-time conversations, that’s not acceptable. Other models like Whisper and Deepgram are much faster, but often gave us unusably bad results for phone audio, especially in non-English languages.
After months of iteration, we've put together a system that glues together a number of AI models, both open-source and API-based, in ways that compensate for each others’ weaknesses while minimizing latency. (Sorry I can’t share more details yet — we’re keeping a few secrets while we try to commercialize this.)
The funny thing is, what we’ve built is probably going to be obsolete soon as foundation models get better and better. But for now, it seems that we’re the first working solution for this problem.
Now that we have a working product, we’ll be exploring B2B use cases for TranslateMyCall.com, but for now and into the foreseeable future, we’re releasing it as a free public alpha. To try it out, go to www.translatemycall.com, select your language and phone number (we support 40 languages and 36 country codes), and the other person’s language and phone number, and both of you will receive a phone call connecting you through an interpreted line.
Please let us know what you think, especially any suggestions for improvement! And if you know someone who might benefit from this, we’d appreciate it if you share it with them. You can reach me with any questions or feedback here or at alex@translatemycall.com.
