Real conversation data for full-duplex speech models
Natural two-speaker English conversations, recorded on isolated, time-aligned tracks, so your model learns the rich interaction of real talk: overlap, backchannels and interruptions.
Talk to an Expert →Get sample pack →What full-duplex data captures
Half-duplex data teaches a model to wait its turn. Real conversation doesn't. These recordings keep everything that happens between turns.
Overlapping speech
Both people talking at once, on separate clean channels, so a model can hear each voice through the overlap.
Overlap eventsPer-speaker tracksBackchannels
The "mm-hmm", "right" and "yeah" that listeners give while someone else holds the floor.
Mm-hmmRight, yeahHesitationsTurn-taking
Interruptions, barge-ins, pauses and the gap between one speaker stopping and the next starting.
InterruptionsTurn gapsBarge-in- Full-duplex speech models
- Voice agents
- Turn-taking & endpointing
- Conversational ASR
- Diarization & separation
- Dialogue speech generation
Explore a sample conversation
One conversation as we deliver it: both isolated tracks, the mix, the transcript and the metadata. Solo either speaker and jump from overlap to overlap.
Listen
Preview audio is compressed; delivered tracks are lossless 48 kHz WAV.Paused · Mixed audio speaker_01 speaker_02 Full Full2m30s 00:00.0 / 0:00 speaker_01speaker_02Overlap 5s 5s Speed 0.75×1×1.25×1.5×2× Skip 5 s10 s30 s Overlap Live transcriptAn end-to-end collection pipeline
One team runs every step from design to delivery, so quality is built in rather than filtered out at the end.
Design
Instructions, speaker targets and protocol shaped around what your model needs.
Recruit & screen
Speakers matched to your profile targets; equipment checked before they are booked.
Record
Hosted remote sessions, each speaker on a lossless isolated track on one clock.
Verify
A trained host monitors every second live, gives a verdict, then automated checks run.
Deliver
Aligned audio, timed transcripts and metadata in a consistent, traceable structure.
Why teams choose Abaka
Heard by people, sourced worldwide, spoken naturally, built to your spec.
Every second, heard by a human
- A live host in every session, listening to both voices as they are recorded
- Fixed on the spot: echo, noise or volume problems are re-recorded during the call
- People decide; automated checks only advise
Voices from around the world
162countries& regions46languages Top regionsUnited StatesIndiaNigeriaKenyaBrazilUnited KingdomEgyptCanadaIndonesiaPhilippines+152 more AccentsAmericanBritishAfricanIndian+ more LanguagesEnglishChineseSpanishArabicHindiFrenchPortugueseRussianJapaneseGermanKoreanIndonesian+34 more
- Tagged per speaker: accent, native language, age and gender in every file's metadata
Real people, real talk
- Real people, not actors or synthetic voices
- Unscripted: each side has different information, so they negotiate, interrupt and react
- Overlaps, backchannels and hesitations kept, not cleaned away
Built to your spec, at scale
SpeakersAccent, native language, age and gender targets InstructionsYour topics and tasks, or ours FormatOne or two speakers · isolated 48 kHz tracks ScaleAdd hosts to add hours, with the same live review- One brief, one team: from design to delivery
Tell us the data
your model needs
Full-duplex model, voice agent or conversational ASR: we'll scope speakers, scenarios and volume with you, and a trained host will hear every second before it reaches you.
Talk to an Expert →Get sample pack →