11 November 202612 min readRelease 01Speech recognitionDarija
Introducing fihrIA-ASR: open-weight Darija speech recognition, better and 7× faster than Voxtral
The engineering story behind the first release: 10,000 hours of transcribed Moroccan Darija audio, an 89k-token tokenizer, a native-speaker-validated metric, and benchmarks from H100 down to a laptop CPU.