Tiro Qwen3-ASR 1.7B (FFASR)
Far-field fine-tune of Qwen/Qwen3-ASR-1.7B-hf, submitted to the Far-Field ASR Leaderboard.
Training
Encoder (audio-tower) and projector only fine tune of Qwen3-ASR-1.7B-hf. Training dataset involves convolution of anechoic sounds from musan-noise and speech from libritts with RIRs from AcousticRooms.
| base model | Qwen/Qwen3-ASR-1.7B-hf |
| trained | audio_tower + multi_modal_projector only |
| speech | LibriTTS train.clean.360, ≥2.0s |
| rooms | AcousticRooms — 260 rooms, 132k RIRs, subsampled from Meta's AcousticRooms |
| noise | MUSAN (bilguun/musan-noise) |
-
Liu, Xiulong, et al. “Hearing Anywhere in Any Environment.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.
-
Mullins, Sarabeth S., Georg Götz, Eric Bezzam, Steven Zheng, and Daniel Gert Nielsen. “Treble10: A High-Quality Dataset for Far-Field Speech Recognition, Dereverberation, and Enhancement.” arXiv preprint arXiv:2510.23141, 2025. Paper
-
Snyder, David, Guoguo Chen, and Daniel Povey. “MUSAN: A Music, Speech, and Noise Corpus.” arXiv preprint arXiv:1510.08484, 2015. Paper
-
Zen, Heiga, et al. “LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech.” Proceedings of Interspeech, 2019, pp. 1526–1530. Paper