AI voices still say your product's name wrong. Building @OnepinAI. Every line checked before it ships. YC. PhD Georgia Tech
SF, Bay Area
- Tweets 521
- Following 117
- Followers 209
- Likes 547
Pinned Tweet
Your AI voice sounds human. So why can't it say your product's name?
A great AI voice reads "Porsche Taycan" as TAY-can. Porsche says TIE-kahn. It guessed from the spelling, and nobody caught it, because nobody listens to line 1,200.
Today we're launching Onepin: the production step after text-to-speech. It checks every line of voiceover before it ships using the voices you already work with.
Onepin can:
➤ Check people's and product names against a 4-million-word pronunciation dictionary
➤ Spell out prices and dates before the voice speaks
➤ Score every line of audio for naturalness, clarity and word accuracy
➤ Fix the one wrong word in the same voice, without re-rendering the take
Works with your voice subscription on @ElevenLabs, @OpenAI, @Google and 30+ more. No phonetic spellings to type. No re-rolls. No switching providers.
Free to start, no credit card required. Hear the before and after in the thread ⬇️
120
68
275
437,861
Over 200k views on this!
Many people have the same problem in AI voice: pronunciation.
People prefers natural + fast response. But, people stops on mispronunciation immediately.
Your AI voice sounds human. So why can't it say your product's name?
A great AI voice reads "Porsche Taycan" as TAY-can. Porsche says TIE-kahn. It guessed from the spelling, and nobody caught it, because nobody listens to line 1,200.
Today we're launching Onepin: the production step after text-to-speech. It checks every line of voiceover before it ships using the voices you already work with.
Onepin can:
➤ Check people's and product names against a 4-million-word pronunciation dictionary
➤ Spell out prices and dates before the voice speaks
➤ Score every line of audio for naturalness, clarity and word accuracy
➤ Fix the one wrong word in the same voice, without re-rendering the take
Works with your voice subscription on @ElevenLabs, @OpenAI, @Google and 30+ more. No phonetic spellings to type. No re-rolls. No switching providers.
Free to start, no credit card required. Hear the before and after in the thread ⬇️
4
12
645
Your AI voice sounds human. So why can't it say your product's name?
A great AI voice reads "Porsche Taycan" as TAY-can. Porsche says TIE-kahn. It guessed from the spelling, and nobody caught it, because nobody listens to line 1,200.
Today we're launching Onepin: the production step after text-to-speech. It checks every line of voiceover before it ships using the voices you already work with.
Onepin can:
➤ Check people's and product names against a 4-million-word pronunciation dictionary
➤ Spell out prices and dates before the voice speaks
➤ Score every line of audio for naturalness, clarity and word accuracy
➤ Fix the one wrong word in the same voice, without re-rendering the take
Works with your voice subscription on @ElevenLabs, @OpenAI, @Google and 30+ more. No phonetic spellings to type. No re-rolls. No switching providers.
Free to start, no credit card required. Hear the before and after in the thread ⬇️
120
68
275
437,861
You can hear the difference: go.onepin.ai/hear-it
1
27
1,412
Same address, two engines.
3042 Saint Luis St NW gets spelled like a math problem. And $1,249.00 still lands as one two four nine dollars.
Leaderboards dont hear street numbers. The miss starts in the normalzier.
onepin.ai
1
6
161
Same changelog, two engines.
v2.4.1-hotfix gets spelled like a phone number. Siobhán still comes out See-oh-ban.
Leaderboards dont hear release notes. Pronuncation is per token.
onepin.ai
1
2
38
Top data stack bookmarks.
@duckdb 41k ⭐ in-process SQL
@dbt_labs Core 14k
@ApacheAirflow 47k
@DataTalksClub Zoomcamp 46k
@metabase 49k
I still query CSVs in DuckDB before I pay a warehouse.
github.com/duckdb/duckdb
1
65
Same product name, three locales.
English stress is off. German compounds snap. Japanese mora timing drifts.
Leaderboards average that away. Voice performence is per language.
onepin.ai
2
26
A new TTS benchmark! Rime released a new benchmark AlphaBench, SupportBench, and ContentBench.
Typical A/B preference benchmark doesn't tell how well your agent will work in the real world.
See: rime.ai/tts-benchmarks
Github: github.com/rimelabs/alpha-tt…
2
40
Amazon blocked @Meta Muse from shopping on Amazon. Muse was the #1 free iOS app last week. @ajassy says the agent never identified itself and capturd credentials.
The storefront is the product. Agents that browse like a person get treated like scrapers.
geekwire.com/2026/amazon-blo…
1
55
Top 12 OSS swaps for paid SaaS.
Coolify instead of Vercel (62k ⭐)
n8n instead of Zapier (205k)
Plane instead of Linear (59k)
NocoDB instead of Airtable (65k)
Supabase instead of Firebase (110k)
I still pay cloud when I dont want to babysit Postgres. The lock-in is optional.
63
Hacktron used @AnthropicAI Opus 5 to break into @OpenAI. $6,500 bounty. Opus 4.8 couldnt write the exploit. Opus 5 did it the next morning.
The model that shipped overnight is the actual patch cadence.
techcrunch.com/2026/09/18/re…
ALT Dario Amodei
60
Arthur Mensch and @MistralAI put Small 4 inside Firefox Smart Window. US, Canada, France first. Zero data retension.
The fight moved from which model is best to which models the brower even lets you pick.
mistral.ai/news/mistral-x-mo…
ALT Arthur Mensch
201
Ursula von der Leyen’s team will propose banning under-15s from social, video, AI chatbots and games.
13-14 only with parents. Age checks on signup.
Chatbots sit next to Instagram in the same draft.
reuters.com/legal/litigation…
ALT Ursula von der Leyen
32
Top growth bookmarks I keep sending founders:
YC Startup School
First Round Review
Lenny's newsleter (1.2M)
@PostHog ⭐️ 39.7k
@plausiblehq ⭐️ 29k
startupschool.org
review.firstround.com
lennysnewsletter.com
2
47
Dario Amodei and @AnthropicAI shipped Claude Fable 5.1. Terminal-Bench-Science 0.1: 52.6% vs GPT-5.6 Sol at 22.4%. Same model as Mythos 5.1, tighter cyber safegaurds.
anthropic.com/claude-fable-a…
ALT Photograph of Dario Amodei
62
Dario Amodei and @AnthropicAI just published the Sep 2026 threat report. Alibaba, Moonshot, DeepSeek: ~200M Claude exchanges harvested for distillation. Qwen campaign hit 151M May–July.
Frontier weights leak through the API, not the checkpoint.
techcrunch.com/2026/09/10/an…
ALT Photograph of Dario Amodei
1
47
Jacob Coxon walked away from @AnthropicAI two months before his equity vested.
Three years of pre-training at @OpenAI and Anthropic. Models that know they are being tested is now a daily fact, not sci-fi.
He left money on the table. That is the signal.
axios.com/2026/09/09/anthrop…
ALT Photograph of Jacob Coxon
37
Course audio fails on names, not on sounding human.
Louis XIV. Siobhán. Yosemite. Same model, three locales, three different misses.
Leaderboards average that away. Voice performence is per language.
onepin.ai
15