Add a benchmark

Benchmarks / Other

VoiceCodeBench

Speech-to-text task success on English workplace recordings that contain exact structured values such as addresses, email, dates, and file paths.

20models
–Open Index weight
2026-09-24last updated
Vals AIpublisher

VoiceCodeBench

Higher is better · best effort per model

Chart settings
20 of 20 selected

full table below

#ModelEffortScoreCost / taskHarness
1GPT Live Transcribenone67.7%$0.019–
2Grok Voice Transcribe 2.0none65.3%$0.0037–
3Cartesia Ink 2none62.0%––
4GPT 4o Transcribenone59.7%$0.0031–
5Deepgram Nova-3 (Streaming)none59.3%$0.0054–
6Reson8 Resonant-1none59.0%$0.0077–
7Inworld STT 1none57.3%$0.0027–
8Whisper Large-v3 Turbonone57.0%$0.0007–
9Muse Voice Transcribenone57.0%$0.0034–
10Deepgram Flux General Englishnone55.7%$0.0072–
11Google Cloud Chirp 3none55.3%$0.018–
12ElevenLabs Scribe v2 Realtimenone55.3%$0.0073–
13Whisper Large-v3none54.3%$0.0021–
14Voxtral Mini Transcribe 2none53.0%$0.0034–
15Grok Voice Transcribe 1.0none52.0%$0.0037–
16GPT 4o Mini Transcribe (Streaming)none51.7%$0.0016–
17Voxtral Mini Transcribe Realtime 2602none46.0%$0.0067–
18Cohere Transcribenone44.7%––
19Azure Speech Universal Language Modelnone37.3%––
20AssemblyAI Universal-3.5 Pronone30.3%$0.0085–