Skip to main content
BenchLM
Data
Official provider results

Liquid AI LFM3 audio and omni evaluation

Liquid AI reports 6.23% English ASR word error rate, 3.83 UTMOSv2 English speech quality and 73.13 on VoiceBench for LFM3-Audio-2B, and 70.09 on Daily-Omni with 46.32 on OmniBench for LFM3-Omni-2B.

Source snapshots refreshed
Five measurement lanesVOICE / S2S

Reasoning, task completion, conversation dynamics, experience, and latency stay separate.

Back to directory

Benchmark profile

Scope
English and multilingual speech recognition, English text to speech, spoken question answering (VoiceBench), audio tool calling, audio understanding (MMAR), and for the omni model, audio-visual video understanding, image-plus-audio understanding and spoken visual question answering
Primary metric
Word and character error rates (%, lower is better); UTMOSv2 on a 1-5 scale and accuracy or benchmark scores (higher is better)
Owner
Liquid AI
Available evidence
Exact values in the Hugging Face model cards and the launch post

What it measures

Spoken reasoning
Does the system understand and answer spoken requests?
Task completion
Can it follow policy, use tools, and finish a workflow?
Voice experience
Is the exchange natural, robust, and responsive?

Available results

Speech and audio

Liquid AI-reported. ASR rows are lower-is-better; the other columns are higher-is-better.

Liquid AI-reported. ASR rows are lower-is-better; the other columns are higher-is-better.
ModelEnglish ASR WERMultilingual ASR WER / CEREnglish TTS UTMOSv2VoiceBenchTool calling (internal)MMAR
LFM3-Audio-2B6.23%6.03%3.8373.1358.5050.60
LFM3-Omni-2B6.36%6.72%3.8064.7352.2051.40

Audio-visual understanding

Liquid AI-reported multiple-choice accuracy. Daily-Omni, WorldSense and AV-SpeakerBench use video with its soundtrack; OmniBench pairs an image with an audio clip, over all 1,142 questions; the speech column is its speech-audio subset.

Liquid AI-reported multiple-choice accuracy. Daily-Omni, WorldSense and AV-SpeakerBench use video with its soundtrack; OmniBench pairs an image with an audio clip, over all 1,142 questions; the speech column is its speech-audio subset.
ModelDaily-OmniWorldSenseAV-SpeakerBenchOmniBench overallOmniBench speech
LFM3-Omni-2B70.0947.1046.7946.3248.25

Spoken visual question answering

Liquid AI-reported. Image benchmarks with each question given only as speech. POPE is F1 over the adversarial, popular and random subsets.

Liquid AI-reported. Image benchmarks with each question given only as speech. POPE is F1 over the adversarial, popular and random subsets.
ModelRealWorldQAMMStarPOPEERQAEmbSpatial
LFM3-Omni-2B64.6851.7983.7839.4367.59

Vision with written questions

Liquid AI-reported. OCRBench is out of 1,000; DocVQA and InfoVQA use ANLS; MMMU-Pro is the vision setting; Video-MME is without subtitles and without audio.

Liquid AI-reported. OCRBench is out of 1,000; DocVQA and InfoVQA use ANLS; MMMU-Pro is the vision setting; Video-MME is without subtitles and without audio.
ModelMMStarMMMU (val)MMMU-ProMMBenchRealWorldQAMathVista (mini)DocVQAInfoVQAOCRBenchRefCOCOMVBenchVideo-MMERoboSpatial-Home
LFM3-Omni-2B55.6742.4427.2873.6866.5462.5092.4764.0880673.5558.0357.7049.14

Interpretation limit

Liquid AI measured every result, including the competitor columns not carried here, with each model’s own chat template and greedy text decoding. English ASR averages the ESB / Open ASR suite; multilingual ASR averages FLEURS over nine languages, with CER for Chinese, Japanese and Korean. VoiceBench is the mean of 9 subsets with the three 1-5 judge scores multiplied by 20. The tool-calling score is an internal evaluation without a public task set. These results do not enter text-model or voice-agent rankings.