abstract background

Speech & audio

The proving ground

for audio intelligence

We test how AI models understand sound, engage in conversations, and perform across languages and dialects.

The hidden infrastructure behind world-class AI models

The hidden infrastructure behind world-class AI models

The hidden infrastructure behind world-class AI models

Featured work

Three benchmarks

Three benchmarks

MedMosaic

Built to find where audio AI breaks, these benchmarks test leading models from Google, OpenAI and others on medical reasoning, live voice agents and unified audio.

  • MedMosaic

    A challenging large-scale benchmark of diverse medical audio.

    Weighted-average accuracy · 13 systems

    GoogleAlibabaOpen / academic
    68.1
    Gemini 2.5 Pro
    60.5
    Gemini 2.5 Flash
    42.8
    Qwen-Omni 7B
    42.1
    Gemma 3n 8B
    41.0
    DeSTA 2.5
    38.6
    Baichuan-Omni
  • DuplexWorld

    Can voice agents help you get through the day?

    Agentic capability · Pass@1 · six worlds

    xAIOpenAIGoogleAmazon
    0.490
    Grok Voice
    0.433
    GPT-Realtime 2.1
    0.398
    Gemini 3.1 Live
    0.188
    GPT-RT 2.1 mini
    0.011
    Nova 2 Sonic
  • TORUS

    A test of rendering–understanding self-coherence.

    Self-coherence · all 432 questions

    Open / academic
    63.2
    Cascade
    50.5
    Audex-30B
    42.8
    Audio-Omni
    35.4
    Audex-2B
    23.4
    UniAudio-2
    21.8
    Unified-IO 2

Research areas

Six areas of speech and audio research.

Six areas of speech and audio research.

Audio reasoning,

From audio reasoning and voice agents to unified audio models, audio safety, multilingual speech and speech enhancement. Select an area to see its papers.

From audio reasoning and voice agents to unified audio models, audio safety, multilingual speech and speech enhancement. Select an area to see its papers.

Audio reasoning

The ability of models to draw inferences from sound, answering questions that depend on what is heard, not only what is said, across speech, environmental sound and music.

MedMosaic · SEABED · CaReCoS · AudioLoop

Voice agents

Spoken-language systems that hold real-time conversations and complete tasks on a user’s behalf, combining speech recognition, reasoning, tool use and speech synthesis.

DuplexWorld

Unified audio models

Single models that both understand and generate audio: transcribing, describing, synthesising and editing speech, sound and music within one architecture.

TORUS

Audio safety

The study of risks in speech and audio systems (bias, misuse, spoofing and harmful outputs) and the methods used to measure and reduce them.

Voice or Stereotype?

Internationalization

Making speech and audio technology work across languages, dialects and accents, so performance holds beyond a handful of high-resource languages.

SEABED · AAE-ICL

Speech enhancement

Techniques that improve the quality and intelligibility of speech by reducing noise, reverberation and distortion in recorded or live audio.

Speech Enhancement from Imperfect Transcriptions

Research papers and publications

Advancing AI Research

Advancing AI Research

Through science and real-world impact.

Through science and real-world impact.

Our research examines the challenges and opportunities shaping AI—from model evaluation and safety to benchmarking and real-world performance.

Our research examines the challenges and opportunities shaping AI—from model evaluation and safety to benchmarking and real-world performance.

SEABED: SouthEast Asian Benchmark for Evaluating Audio Reasoning

Sep 18, 2026

Voice or Stereotype? Disentangling Acoustic and Content-Based Gender in Speech-to-Speech Models

Sep 8, 2026

DuplexWorld: Can voice agents help you get through the day?

Aug 11, 2026

TORUS: A Test of Rendering-Understanding Self-Coherence for Unified Audio Models

Jul 30, 2026

Connect with Centific

Stay ahead of what’s next

Stay ahead

Updates from the frontier of AI data.

Receive updates on platform improvements, new workflows, evaluation capabilities, data quality enhancements, and best practices for enterprise AI teams.

By proceeding, you agree to our Terms of Use and Privacy Policy