ASR with timestamps from commercial speech APIs
| System | Mean F1 |
|---|---|
| Google Chirp 2 | 0.486 |
| Google Chirp 3 | 0.452 |
| Qwen3-ASR-Flash | 0.363 |
| Mistral Voxtral Mini | 0.308 |
| ElevenLabs Scribe v2 | 0.281 |
Word timestamp F1 at 20 ms, mean of TIMIT core test and Buckeye test, each clean and with reverb, noise, music and babble added. Higher is better.
Olewave makes FA-Bench. It scores every system with the same public code against the same human annotations. Olign, also from Olewave, no longer takes part in the benchmark.
We proudly offer
Large-Scale Pre-Labeled Speech Datasets
Label-Validated,
and Scientist-Verified OTS Datasets
-
Multiple languages, including 🇺🇸 🇪🇸 🇲🇽 🇸🇦 🇧🇷 🇮🇳 🇯🇵 🇨🇳 🇬🇧 🇩🇪 🇫🇷 …
-
Diverse topics in education, finance, legal, entertainment, healthcare, retail, customer service …
-
Validated labels for speaker names & turns, transcripts …
Production-Grade Voice AI Solutions
On Your Own Servers,
No SaaS Vendor Lock-In.
-
Superior STT/TTS with higher quality & faster inference than open-source or cloud APIs
-
Tailored to your use case, tuned to your domain, languages, and workflows
-
You own the code, with 100% IP rights on all customization
Avant-Garde Voice AI Consulting Services
Experienced in Assisting Clients,
Located in San Francisco
-
Market & technology evaluation covering the competitive landscape and build-vs-buy
-
Strategic business planning for roadmap, staffing and go-to-market
-
Strict non-disclosure, with an NDA on every engagement