Backed by Y Combinator

We make AI understand you.

Leading data research lab on human interaction powering the next generation of speech and omni-models.

Our Mission

Diverse and complex human interactions enabled us to land on the moon. Interaction is much more than handling end-of-turns and barge-ins.

Our mission is to enable the most emotionally intelligent AI models that unlock the next level of human-AI interaction.

What are the best conversations you've had recently? That's the bar we're aiming for.

Our Research

We publish original research and open source our datasets on production failures we see in voice systems today.

EvaluationSeptember 10, 2026

Duplex Cue

An audio benchmark for measuring how an ongoing speaker responds when a listener contributes during the turn—whether the speaker continues unchanged, adapts in-turn, or yields the floor.

Duplex Cue assigns two independent labels to each local interaction. Cue intent describes Speaker B's contribution. Observed response describes Speaker A's behavior.

Fine-tuningAugust 28, 2026

Inkling Post-Training

We post-trained Thinking Machines' Inkling on selections of 1, 25, and 100 hours of VoiceCodeBench-style speech.

Our training goal was not to teach Inkling a lookup table of canonical values. It was to make the model better at hearing the acoustic structure that makes those values recoverable.

BenchmarkJuly 7, 2026

Vocal Affect Bench

A public, test-only benchmark for evaluating whether AI audio models can identify expressed vocal emotion from raw speech — without transcripts or contextual metadata.

A customer who says "that's fine" may be conceding, joking, or signaling frustration. Automatic speech recognition preserves the phrase but strips the vocal delivery that makes the exchange interpretable.

BenchmarkMay 1, 2026

Voice Code Bench

WER measures transcript similarity. Production voice interfaces need something stricter: the exact email, command, account identifier, URL, date, file path, dollar amount, product code, or measurement a downstream system must act on.

Voice Code Bench is intentionally narrower than universal audio suites. It asks one product-critical question: did the transcript preserve every structured entity a downstream workflow needs?

EvaluationSeptember 10, 2026

Duplex Cue

An audio benchmark for measuring how an ongoing speaker responds when a listener contributes during the turn—whether the speaker continues unchanged, adapts in-turn, or yields the floor.

Duplex Cue assigns two independent labels to each local interaction. Cue intent describes Speaker B's contribution. Observed response describes Speaker A's behavior.

Fine-tuningAugust 28, 2026

Inkling Post-Training

We post-trained Thinking Machines' Inkling on selections of 1, 25, and 100 hours of VoiceCodeBench-style speech.

Our training goal was not to teach Inkling a lookup table of canonical values. It was to make the model better at hearing the acoustic structure that makes those values recoverable.

BenchmarkJuly 7, 2026

Vocal Affect Bench

A public, test-only benchmark for evaluating whether AI audio models can identify expressed vocal emotion from raw speech — without transcripts or contextual metadata.

BenchmarkMay 1, 2026

Voice Code Bench

Voice Code Bench is intentionally narrower than universal audio suites. It asks one product-critical question: did the transcript preserve every structured entity a downstream workflow needs?

EvaluationSeptember 10, 2026

Duplex Cue

An audio benchmark for measuring how an ongoing speaker responds when a listener contributes during the turn—whether the speaker continues unchanged, adapts in-turn, or yields the floor.

Duplex Cue assigns two independent labels to each local interaction. Cue intent describes Speaker B's contribution. Observed response describes Speaker A's behavior.

Fine-tuningAugust 28, 2026

Inkling Post-Training

We post-trained Thinking Machines' Inkling on selections of 1, 25, and 100 hours of VoiceCodeBench-style speech.

Our training goal was not to teach Inkling a lookup table of canonical values. It was to make the model better at hearing the acoustic structure that makes those values recoverable.

BenchmarkJuly 7, 2026

Vocal Affect Bench

A public, test-only benchmark for evaluating whether AI audio models can identify expressed vocal emotion from raw speech — without transcripts or contextual metadata.

BenchmarkMay 1, 2026

Voice Code Bench

Voice Code Bench is intentionally narrower than universal audio suites. It asks one product-critical question: did the transcript preserve every structured entity a downstream workflow needs?

Our Data

From everyday conversations to specialist domains, built around the way people actually speak.

Browse inventory

Conversation

Long conversations

Natural, extended, full-duplex conversation data between two speakers.

Training environments

Reinforcement learning

Environments built using real human speech to improve instruction following and conversational handling.

Languages & accents

Multi-lingual

Speech across diverse languages and accents, spanning more than 100 countries and locales.

Group dynamics

Multi-speaker

Natural conversations between three or more speakers.

Domain knowledge

Specialist domains

Conversations or script reading in finance, medicine, education, and other specialist fields.

Professional voices

Voice actors

Professional voice actor recordings with a custom commercial license.

Have interesting data you can license? Get in touch with us!

Our Approach

We understand the details of human interactions so we can train models too.

turn-taking: Know when to speak, when to listen, and when someone still has more to say.
backchannels: Offer a well-timed “mm-hmm” or “right” without taking over the conversation.
empathetic responses: Respond with words and vocal delivery that acknowledge the speaker’s situation.
noisy speech: Follow the intended speaker through traffic, music, and everyday background noise.
accent robustness: Understand people across accents without asking them to change how they speak.
emotional control: Shape the tone and intensity of a response to suit the context.
interruption recovery: Yield when someone cuts in, understand the change, and pick up the right thread.
multi-speaker interaction: Follow a conversation as several people contribute, overlap, and exchange turns.
prosody: Use pitch, stress, and intonation to carry meaning beyond the words.
code-switching: Follow people as they switch languages, even within the same sentence.
active listening: Show understanding through relevant acknowledgments, questions, and responses.
full-duplex listening: Keep listening while speaking, and adapt as the conversation changes.
conversational repair: Recover from a misunderstanding and get the conversation back on track.
context continuity: Carry the relevant details forward across turns, interruptions, and topic changes.
speaker tracking: Keep track of who said what as voices and turns change.
pacing & rhythm: Vary speech rate and rhythm to make explanations, reassurance, and dialogue feel natural.
vocal nuance: Attend to subtle changes in emphasis, delivery, and vocal expression.
natural pauses: Leave room for a breath, a thought, or an unfinished sentence.
response timing: Respond at a pace that fits the moment, without rushing or leaving awkward gaps.
overlap handling: Distinguish a supportive interjection from an attempt to take the floor.
emotion recognition: Notice expressed emotion in the voice, beyond what the transcript alone conveys.
laughter & sighs: Understand and express the nonverbal sounds that add meaning to conversation.
tone adaptation: Adjust vocal delivery as the conversation shifts from playful to serious, or uncertain to clear.
dialect coverage: Handle regional pronunciation, vocabulary, and ways of expressing an idea.
far-field speech: Understand speech captured at a distance, with room echo and less direct sound.
disfluency handling: Make sense of hesitations, repetitions, false starts, and mid-sentence corrections.
environmental context: Use relevant sounds around the speaker to interpret the situation.
addressee awareness: Recognize when speech is directed at the model or at someone else nearby.
clarifying questions: Ask a useful follow-up when a request is unclear or something was misheard.
intent understanding: Understand what someone is trying to communicate, including indirect requests.

Our Leaderboards

Benchmarks built with real human speech for real production failures.

Transcription / ASR

Voice Code Bench

How well are models able to transcribe key entities such as phone number and emails? Given a list of critical entities, we calculate the Task Success Rate, which is the % of recordings where every target entity is recovered correctly.

Task Success RateHigher is better

Loading results…

Emotion Detection

Vocal Affect Bench

How accurately can models identify emotions expressed in speech? We measure accuracy as the % of scored clips where the expressed emotion is identified correctly, across seven emotion categories.

AccuracyHigher is better

Loading results…