AI voices sound human now. But they still guess names from the spelling, and nobody has time to listen to every line. Onepin runs on the voices you already use (ElevenLabs, OpenAI, Google, 30+ more), spells out prices and dates, checks names against a 4M-word pronunciation dictionary, scores every line, and fixes the wrong word without re-rendering the take. Built for teams shipping AI voiceover in several languages: product videos, dubbing, courses. Free to start, no credit card required.
Free Options
Launch Team / Built With
AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
@rukhsar_amjad so you have an audio? it is something we are making now. You put the audio and the script, and we automatically fix and return the audio. Stay tuned. Please sign up, you will receive product updates.
@jiwon_hong If you see an ad saying their product name wrong, you won't buy. It is a deal breaker.
Interesting project, I think you nailed a good problem to solve.
@shawn_cao Thanks. Stay tuned. We have more interesting features coming out soon.
A lot of the names I have to get right are people and local places no dictionary carries. What happens to a word that isn't in the four million?
The drug terms in the 4M-word dictionary caught my eye. I'm building a voice AI that phones elderly parents every day, and the failures that hurt trust most are exactly this kind: a medication name or a family member's name said slightly wrong, and an older listener stops believing the call is from someone who knows them. Question: does the pronunciation check work on short, live-generated lines (a few seconds, no time to re-render), or is it designed for batch scripts like videos and courses? And can a user pin a custom name, like a grandchild's, so it is always said right?
@igorgurovich Great question: for a full transparency, we have been focusing on the accuracy while sacrificing the latency. Our roadmap is to make AI accurate first and then to make them fast. I have been hearing from numerous voice AI agent companies that the errors on critical terms are deal breakers, we are working toward making them possible. Technically, it is called a distillation (or teacher-student network) where a big neural network teaches a small one that typically runs fast, we are on this. Please contact me directly.
Getting names, prices and dates pronounced correctly is the small detail that makes AI voiceovers usable in real work, since a garbled client name is an instant give away. Checking against a four million word pronunciation dictionary and scoring each line is a thorough way to catch the misses. Does it let you save the corrections so the same name is right next time?
@karimbenkeroum Oh great idea. We spent months for building 4M dictionary, but it doesn't cover all the new words. So our agent instantly looks for pronunciations from web, YouTube, etc, and uses them for correct generation, and most importantly remembers them. What is your use case?
the naming problem is real, we deal with the same thing on live calls at Dial. the difference for us is latency, we can't check and re-render before the line ships because there's no "before" at all, it's a live turn. is Onepin scoped to pre-rendered scripts where you have time to check first, or could the pronunciation dictionary and scoring run fast enough to sit in a live TTS path too
@galdayan great question. I understand your concern: on the voice AI agent call, you want to make it work. It is on our roadmap. We understand your problem, we decided to first solve the error issue first and then address the speed issue. Small hint: distillation. Our small network (under training) will learn the big network that is currently working as a part of the current Onepin service. Stay tuned!
@soohyun distillation makes sense as the path, that's basically the only way to keep accuracy while cutting inference time without retraining from scratch. curious whether the student model will still get to hit your 4M-word dictionary lookup at runtime, or whether that lookup itself becomes part of what gets distilled into the weights. if it stays an external lookup the latency math probably still works for a live call.
@galdayan Good call. The student model may remember the dictionary, or needs MoE (mixture of student models per locale). We don't know yet. Please sign up, you will receive emails about the product update.
@galdayan You are an expert. There are several technical tradeoffs, we will carefully deal with them. Thank you!