VoiceArena has opened its Speech Recognition Benchmark with the Monsoon Conversational ASR Test Set, ranking speech-to-text models on two-person phone conversations in US English, Indian English, Brazilian Portuguese, Romanian, and Hindi. The live boards score 41 models on US English, 39 on Indian English, 37 on Brazilian Portuguese, 27 on Romanian, and 24 on Hindi. Labs and product teams can use the results for a third-party comparison on conversational audio, the kind of speech deployed voice systems actually have to transcribe. Every clip is a real phone call, recorded in-country on everyday phones, with room noise and background sound kept in. Public splits are on Hugging Face so teams can inspect the audio and transcripts themselves. Matching private splits stay held out, which lowers the chance that a model has already trained on the test set. VoiceArena reports that public and private audio quality match within 0.03 MOS in every language. This is the first dataset in a wider benchmark. Co-founders Shobhit Banga and Supriya Paul, previously of Josh Talks, say more sets will follow over the coming days, covering customer-service audio, dictation, child speech, long-form speech, and far-field speech, with a stated target of at least 100 languages. Where a single word error rate is a weak fit, the design calls for other scores, including lattice-style metrics that accept more than one valid transcript.