Quebec's API dashboard displays precise word timestamps generated from uploaded speech audio.
*Yarn's new Quebec API promises sub‑20 ms word boundary precision, a claim that could reshape video editing pipelines. The startup pits its custom decoder and Whisper‑based CNN against established forced aligners, offering a free trial before moving to a modest pay‑as‑you‑go model.*
Yarn, a W24‑stage startup known for its video‑editing platform, unveiled Quebec, an API that promises word‑level timestamps accurate to within 20 ms. In the world of talking‑head video, a 50 ms mis‑cut can ruin a take, so precision matters. Existing forced aligners—Montreal Forced Aligner and ElevenLabs Cue—failed Yarn's standards, prompting a home‑grown solution. The company released a limited free trial, urging developers to test the claims before a modest hourly fee kicks in.
Yarn built Quebec around Montreal Forced Aligner's acoustic model but stripped away the costly beam search. Their decoder forces each word to land within 200 ms of a rough estimate, shrinking the search space enough for an exact solution. On top of that they added a tiny CNN derived from the first two layers of Whisper Tiny. The network nudges each boundary by up to 50 ms, tightening the final timestamps. Lacking commercial‑grade hand‑labelled data, Yarn used Gemini to annotate 17 k word boundaries in a few hours for under $200. The resulting model matched human‑label performance, scoring 73.9 % versus 74 % on the same validation set.
Accuracy is measured as the percentage of word boundaries that fall within 20 ms of hand‑labelled references. On the Buckeye corpus—American speech not seen during training—Quebec hit 75 %. By comparison, the Montreal Forced Aligner managed 61 % and ElevenLabs Cue lagged at 30 %. Those figures suggest Quebec delivers a tangible edge for editors who cannot tolerate even a 50 ms slip in talking‑head footage.
The numbers speak plainly: Quebec outperforms both MFA and ElevenLabs Cue on the same benchmark. Its custom decoder eliminates the need for beam search, while the Whisper‑based CNN provides fine‑grained adjustments. The trade‑off is speed; the service is slower than ElevenLabs. Still, for workflows where precision outweighs latency, Quebec offers a clear advantage. Yarn invites users to test the API for free until tomorrow night, then switch to a pay‑per‑hour plan.
After the free window, Quebec charges $0.32 per hour of processed audio, with a minimum $10 top‑up and no signup required. Yarn serves the service at cost. The API currently supports only English and runs slower than ElevenLabs Cue. Users can submit raw audio or pair it with their own transcript via the docs at https://quebec.yarn.so. Yarn asks for feedback on poor samples and welcomes superior aligners, indicating a willingness to iterate.
Quebec is Yarn's API that returns word‑level timestamps from speech, using a custom decoder around MFA's acoustic model and a Whisper‑Tiny‑based CNN.
On the Buckeye corpus, Quebec placed 75% of word boundaries within 20 ms of hand labels, while MFA achieved 61%.
After the free period, Quebec costs $0.32 per hour of audio, with a minimum $10 top‑up and no signup required.
Quebec's early results position it as a serious challenger to entrenched aligners, especially for creators who cannot afford timing errors. While its English‑only scope and slower throughput temper enthusiasm, the transparent pricing and open invitation for feedback suggest Yarn will keep sharpening the tool. The next wave of video editing pipelines may soon rely on sub‑20 ms word alignment as a baseline, not a luxury.
Sources: Hacker News post – Show HN: An API for word timestamps from speech, 75% within 20ms