tiro-qwen3-asr-1.7b-v2

Tiro is a far-field fine-tune of Qwen3-ASR-1.7B, with better performance on reverberant and noisy speech, scaling as conditions get harder, while maintaining general ASR capabilities. See benchmarks below for more details.

Currently #1 on the FFASR leaderboard.

English only, bf16. Same architecture and parameter count as the base.

Benchmarks

Lower is better. ✅ marks the better result.

FFASR Leaderboard

Condition	Base (WER)	Tiro (WER)
Average	13.41	12.01 ✅
Near field speech	3.76 ✅	3.80
High SNR	6.74	6.22 ✅
Mid SNR	13.89	12.00 ✅
Low SNR	29.26	26.02 ✅

Other benchmarks

Benchmark	Condition	Base (WER)	Tiro (WER)
VOiCES	real rooms, replayed speech	8.88	8.30 ✅
CHiME-4	real noisy environments	4.71	4.19 ✅
CHiME-4	simulated arm	7.00	6.63 ✅
VITW-Bench	8 perturbations, English	7.09	6.40 ✅
NOIZEUS	additive noise, 0-15 dB	9.23	7.85 ✅
Treble10-Speech	reverberation only	3.29	3.19 ✅
AMI	distant mic (SDM)	19.73	19.33 ✅
AMI	headset control (IHM)	8.24 ✅	8.28
Open ASR Leaderboard	mean of 8 public English splits	4.311	4.310

Per-condition breakdown

Benchmark	Condition	Base (WER)	Tiro (WER)
VOiCES	near microphone	3.09	2.85 ✅
VOiCES	far microphone	14.68	13.76 ✅
VOiCES	no distractor	3.27	3.07 ✅
VOiCES	music	8.06	7.08 ✅
VOiCES	television	8.88	8.29 ✅
VOiCES	babble	15.33	14.77 ✅
CHiME-4 real	bus	5.98	5.42 ✅
CHiME-4 real	cafe	4.84	4.09 ✅
CHiME-4 real	pedestrian	4.18	3.89 ✅
CHiME-4 real	street	3.82	3.35 ✅
NOIZEUS	0 dB SNR	23.24	20.76 ✅
NOIZEUS	5 dB SNR	8.42	6.51 ✅
NOIZEUS	10 dB SNR	3.20	2.74 ✅
NOIZEUS	15 dB SNR	2.07	1.39 ✅
VITW-Bench	real recordings	5.30	4.84 ✅
VITW-Bench	simulated	7.80	7.03 ✅

Open ASR Leaderboard

Base column is the published Qwen3-ASR-1.7B result on the 8 public English splits. The leaderboard's headline average also includes two private sets that are not publicly runnable, so the figure shown there is higher than this one.

Split	Base (WER)	Tiro (WER)
LibriSpeech test-clean	1.26	1.25 ✅
LibriSpeech test-other	2.94 ✅	2.95
VoxPopuli	2.85	2.62 ✅
AMI	8.31 ✅	8.45
GigaSpeech	7.22 ✅	7.23
Earnings22	5.84	5.70 ✅
SPGISpeech	2.57 ✅	2.77
Monsoon	3.50 ✅	3.51
mean	4.311	4.310

Usage

Identical to the base model.

from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration

model_id = "rohansheth/tiro-qwen3-asr-1.7b-v2"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3ASRForConditionalGeneration.from_pretrained(model_id, device_map="auto")

Citations

Trained by KL distillation from the frozen base model on noise-free audio, following:

@inproceedings{li2017large,
  title     = {Large-Scale Domain Adaptation via Teacher-Student Learning},
  author    = {Li, Jinyu and Seltzer, Michael L. and Wang, Xi and Zhao, Rui and Gong, Yifan},
  booktitle = {Interspeech},
  year      = {2017},
  eprint    = {1708.05466}
}

@inproceedings{mosner2019improving,
  title     = {Improving Noise Robustness of Automatic Speech Recognition via
               Parallel Data and Teacher-Student Learning},
  author    = {Mo{\v{s}}ner, Ladislav and Wu, Minhua and Raju, Anirudh and
               Parthasarathi, Sree Hari Krishnan and Kumatani, Kenichi and
               Sundaram, Shiva and Maas, Roland and Hoffmeister, Bj{\"o}rn},
  booktitle = {ICASSP},
  year      = {2019},
  eprint    = {1901.02348}
}

Base model:

@misc{qwen3asr,
  title  = {Qwen3-ASR},
  author = {Qwen Team},
  year   = {2026},
  url    = {https://huggingface.co/Qwen/Qwen3-ASR-1.7B-hf}
}