Modulate Tops Hugging Face's Transcription Benchmark

Modulate, a conversational voice intelligence company, now ranks #1 on Hugging Face’s Open ASR Leaderboard, a public benchmark for automatic speech recognition.

The milestone underscores how Modulates voice-native architecture can outperform much larger players in accuracy, speed, and cost.

"Transcription has become foundational to voice AI, but the economics have not kept up with how these systems are actually being deployed," said Mike Pappas, CEO and co-founder of Modulate, in a statement. "Developers and enterprises should not have to choose between accuracy, speed, and affordability. Modulate delivers all three, while opening the door to a much deeper understanding of what is happening in live conversations."

The Hugging Face Open ASR Leaderboard compared leading open-source and commercial transcription models across standardized datasets spanning multiple domains, accents, and recording conditions. Models are evaluated using word error rate. Modulate ranked #1 out of 88 models.

Modulate trains its models on more than 500 million hours of noisy, real-world audio. The model transcribes faster than real time.

Transcription is part of Modulate's broader voice intelligence platform. Modulate's Ensemble Listening Model, or ELM, architecture combines dozens of audio-native models to understand voice.

In addition to high-accuracy transcription, Modulate's transcription models support advanced voice intelligence capabilities, including emotion detection, diarization, accent identification, deepfake detection, and support for nearly 60 languages and dialects.

"Transcription is an important starting point, but it is not the end state," Pappas said. "The real opportunity is conversation understanding. Voice carries signals like emotion, urgency, hesitation, accent, identity, and authenticity that never appear in a transcript. Velma is built to help enterprises capture those signals and turn them into actionable intelligence."

Hugging Face's Open ASR Leaderboard evaluates models across seven datasets, including AMI, Earnings-22, GigaSpeech, LibriSpeech Clean, LibriSpeech Other, SPGI Speech, and VoxPopuli-AA-Cleaned.