List your company

Speech Recognition on Switchboard

6.6WERAR Beam
2.51230.10657.785.294Dec 6, 2022Jul 8, 2023Feb 8, 2024Sep 10, 2024Apr 12, 2025Nov 13, 2025Jun 16, 2026

Evaluation results

MethodLinks
AR Beam2026.066.6———
AR Greedy2026.066.9———
Audio Flamingo 3 + CoATBackbone=Audio Flamingo 3, Method=CoAT2026.067.18———
NAR-MBR|Z|=64, Niter=12026.067.3———
NAR-MBR|Z|=256, Niter=12026.067.3———
NAR-MBR|Z|=256, Niter=102026.067.3———
Qwen2.5-Omni-7B + CoATBackbone=Qwen2.5-Omni-7B, Method=CoAT2026.067.35———
NAR-MBR|Z|=64, Niter=102026.067.4———
NARNiter=102026.067.6———
NAR-MBR|Z|=256, Niter=02026.067.7———
NARNiter=12026.067.8———
NAR-MBR|Z|=64, Niter=02026.067.8———
NARNiter=02026.067.9———
Audio Flamingo 3Backbone=Audio Flamingo 3, Method=Baseline2026.068.01———
Qwen2-Audio + CoATBackbone=Qwen2-Audio, Method=CoAT2026.069.25———
ECLMCorrection model size=0.5B, Decoding strategy=greedy, Base model=Parakeet-TDT-v22024.0510.3——7
Parakeet-TDT-v2Correction method=None2024.0511.1———
Seed-ASR (ML)Language=EN2024.0711.59———
WhisperBackbone=Large V2, Evaluation Protocol=Zero-shot2022.1213.851.2——
Whisper Large-v2Language=EN2024.0713.8———
Qwen2.5-Omni-7BBackbone=Qwen2.5-Omni-7B, Method=Baseline2026.0614.83———
KM 25HzToken Rate=252026.0115—9.5—
KM 12.5HzToken Rate=12.52026.0117.3—11.2—
Kanade 25HzToken Rate=252026.0118.6—11.7—
Qwen2-AudioBackbone=Qwen2-Audio, Method=Baseline2026.0619.06———
Kanade 12.5HzToken Rate=12.52026.0124.6—15.9—
FACodecToken Rate=4802026.0125.5—16.4—
DualCodecToken Rate=1002026.0128.2—18.1—
wav2vec 2.0Backbone=Large, Language Model=None2022.1228.3———
PASTToken Rate=4002026.0128.9—18.8—
TiCodecToken Rate=3002026.0129.1—18.9—
STToken Rate=4002026.0129.4—19.1—
MimiToken Rate=1002026.0130.6—20.1—
StableCodecToken Rate=252026.0145—30.2—
WavTokenizerToken Rate=402026.0167.2—46.8—
X-Codec 2Token Rate=502026.01103.3—75.6—
BiCodecToken Rate=502026.01108.8—78.4—