Far-Field ASR Leaderboard: Treble and Hugging Face Launch FFASR

We are genuinely excited to share, in partnership with Treble Technologies and Hugging Face, the launch of the Far Field ASR (FFASR) Leaderboard: the first open, community-driven far-field ASR leaderboard built to measure how speech recognition models actually perform in real-world acoustic conditions. This is exactly the kind of work our community cares about, and we are proud to be a home for the announcement.

The leaderboard is live on Hugging Face: https://huggingface.co/spaces/treble-technologies/ffasr

What is the Far-Field ASR Leaderboard?

The FFASR Leaderboard is an open speech recognition benchmark that ranks ASR models on how they hold up in far-field conditions rather than the clean lab setup most models are quietly tuned for. Developers, researchers, and enterprise teams can upload their ASR models and see performance across realistic scenarios: reverberation, background noise, competing speech, microphone distance, and rooms with very different acoustics. It has already drawn interest from model builders at NVIDIA, IBM, Cohere, and others.

It builds on the existing Treble and Hugging Face collaboration, which previously produced the openly available Treble10 far-field dataset. FFASR is the natural next step: a public scoreboard that moves audio AI evaluation beyond near-field assumptions.

Why far-field speech recognition is so hard to benchmark

Voice interfaces now live in phones, watches, earbuds, voice agents, humanoids, and cars, so real-world reliability has become critical. The catch is that building physical labs and collecting recordings for every possible environment is slow, expensive, and at scale almost impossible. So most teams evaluate ASR models on clean, close-mic speech with minimal far-field testing. Models score beautifully, then meet reverberant meeting rooms, noisy public spaces, and overlapping speech, where word error rates can climb fast.

Classic far-field benchmarks like CHiME and AMI helped surface this gap, but the industry has lacked an open, standardized, easy-to-enter way to measure it at scale. That is the hole FFASR is built to fill, and why we think this is such a strong launch.

How Treble simulates real-world acoustics

Treble's acoustic simulation is what makes the benchmark possible. Its cloud-based, physics-based engine recreates far-field conditions virtually, so an evaluation set can be as varied and complex as the real world without renting rooms or running microphones. The hard, costly part, recreating thousands of acoustic environments, becomes a simulation problem instead of a logistics one. For the first time, that gives the industry a practical way to measure real-world ASR performance at scale.

What Treble, Hugging Face, IBM, and Cohere are saying

The people building this say it best.

Dr. Finnur Pind, CEO and co-founder of Treble Technologies:

"The speech recognition industry has lacked a non-proprietary, community-driven way to measure how models perform outside ideal laboratory conditions. The Far Field ASR Leaderboard demonstrates how the Treble approach can help developers now evaluate models against the kinds of acoustic challenges users encounter every day. By partnering with Hugging Face, we are making realistic, transparent evaluation accessible to the broader speech AI ecosystem."

Eric Bezzam, Audio ML Engineer at Hugging Face:

"As voice interfaces expand into smart glasses, robotics, and other hands-free applications, evaluating ASR performance in noisy and far-field environments becomes increasingly important. The FFASR Leaderboard is a significant step toward real-world evaluation."

Dr. George Saon, Manager of Speech Technologies at IBM Research, makes the case that this matters even for the giants: automatic speech recognition is not a solved problem, and FFASR is a helpful tool for measuring progress in challenging acoustic environments.

And from a builder's seat, Cohere's Julian Mack, Member of Technical Staff on the Foundations team, notes how useful it is to see dry-speech word error rates next to far-field ones, because it separates core recognition quality from the extra difficulty real acoustics introduce.

Why we are proud to share this on Voice AI Space

Open, transparent benchmarks are how this field moves forward, and far-field performance has been one of its longest-running blind spots. A launch that brings Treble, Hugging Face, and developers from across the ecosystem together to fix it is exactly the kind of story we built this newsroom for. If you are tracking who is shaping voice AI, this one belongs on your radar.

How to join the FFASR webinar and submit your model

The leaderboard is live on Hugging Face, so you can upload your ASR model and see where it lands. Treble and Hugging Face are also hosting a live webinar on Thursday, June 11, 2026, covering how to navigate the leaderboard and submit models, the physics behind the high-fidelity simulation data, and a live Q&A.

Speakers:

Eric Bezzam, Audio ML Engineer, Hugging Face
Dr. Daniel Gert Nielsen, Senior Product Manager for the Treble SDK
Professor Shinji Watanabe, Carnegie Mellon University
Nithin Rao Koluguri, Senior Research Scientist, NVIDIA
Julian Mack, Member of Technical Staff, Foundations, Cohere
Dr. George Saon, Manager, Speech Technologies, IBM Research

Session times:

Europe: 9:00 AM UTC / 10:00 AM CET
USA: 9:00 AM PDT / 12:00 PM EDT / 4:00 PM UTC

Explore Treble's work on Hugging Face: https://huggingface.co/treble-technologies

Published in partnership with Treble and Hugging Face.