Voice & Conversational Interfaces benchmarks
How performance is measured at the frontier · as of 2026-10-26 · 5 benchmarks
Accuracy
Word Error Rate
% WERTranscription errors across the Open ASR test suite — lower is better.
↓ lower is better·Open ASR Leaderboard
- #012.8%
- #023.0%
- #03
microsoft/azure-speech-07-2026 3.2% - #043.2%
- #053.3%
- #06
ibm-granite/granite-speech-5.0-2b 3.6% - #07
TheStageAI/thewhisper-large-v3-turbo 3.8% - #083.8%
- #09
ibm-granite/granite-speech-5.0-470m-turboctc-nc 3.9% - #10
nvidia/parakeet-tdt-0.6b-v2 (fast-gpu-asr) 3.9% - #113.9%
- #12
nvidia/parakeet-tdt-0.6b-v3 (fast-gpu-asr) 3.9%
Speed
Throughput (RTFx)
× real-timeAudio-seconds transcribed per second of compute — higher is faster.
↑ higher is better·Open ASR Leaderboard
- #01
soundsgoodai/Zipformer-cr-ctc-transducer-XL-290M (fast-gpu-asr, ctc_greedy_search) 24,148× - #02
soundsgoodai/Zipformer-cr-ctc-transducer-XL-290M (fast-gpu-asr, transducer_modified_beam_search) 21,814× - #03
ibm-granite/granite-speech-5.0-470m-turboctc 21,134× - #04
ibm-granite/granite-speech-5.0-470m-turboctc-nc 20,801× - #05
nvidia/parakeet-tdt-0.6b-v2 (fast-gpu-asr) 16,943× - #06
nvidia/parakeet-tdt-0.6b-v3 (fast-gpu-asr) 14,126× - #07
abr-ai/niagara-9m-batch.en 7,217× - #08
nvidia/parakeet-tdt-0.6b-v3 6,228× - #09
nvidia/parakeet-tdt-0.6b-v2 6,148× - #105,137×
- #11
abr-ai/niagara-38m-batch.en 4,561× - #12
abr-ai/niagara-19m-batch.en 4,348×
Naturalness
- #01
Audio8-TTS-Realistic-0.6B 1,726 - #021,610
- #031,560
- #041,552
- #051,549
- #061,547
- #071,542
- #081,538
- #091,530
- #101,530
- #111,529
- #121,523
Economics
Transcription cost
$ / minutePublished speech-to-text API price per minute — lower is cheaper.
↓ lower is better·LiteLLM (open pricing)
- #01$0.0007
- #02$0.0020
- #03$0.0037
- #04$0.0043
- #05$0.0060
- #06$0.0060
TTS cost
$ / 1M charsPublished text-to-speech price per million characters — lower is cheaper.
↓ lower is better·LiteLLM (open pricing)
- #01$4
- #02$15
- #03$16
- #04$30
- #05$30
- #06$60
- #07$80