P1 Chilly
isolated voice channel
Voice Research Company
Research-grade voice data for AI labs.
Frisson Labs studies live multiplayer conversation and turns fully consented sessions into channel-separated voice datasets for voice-to-voice model training, evaluation, and research.
Company Thesis
We study voice in live human interaction.
Voice-to-voice models need more than clean read speech. They need real turn-taking, interruptions, hesitation, laughter, overlapping speakers, noisy environments, team callouts, and response timing.
Frisson Labs turns fully consented multiplayer sessions into high-quality duplex voice corpora: channel-separated audio, labeled conversational events, transcripts, participant metadata, rights documentation, and delivery formats that model teams can train on directly.
Audio Sample
Inspect separated speaker channels.
This call segment shows the buyer experience: monitor a mixed call, solo individual speakers, inspect overlap-heavy speaker activity, and see the voice data package around one session.
Session FL-VOICE-042
00:00 / 00:25
P2 king_calvin
isolated voice channel
P3 Gordo
isolated voice channel
P4 Hapuum
isolated voice channel
Demo audio data card
sample_rate_hz
48000
channels
4 speakers + mix
conversation_type
duplex team call
labels
diarization, overlap, speech act
consent_scope
internal demo sample
Speaker activity and overlap markers
Audio Quality
What matters for duplex voice models.
Each participant is delivered as an isolated track, plus a mixed reference monitor for reconstruction and evaluation.
Collections can be scoped for 48 kHz audio delivery, with codec/container choices agreed during licensing.
Sessions include interruptions, barge-ins, backchannels, fast callouts, and multi-speaker overlap windows.
Deliveries can include diarization, speech acts, emotion/reaction tags, SNR checks, clipping flags, and train/eval splits.
Voice Data Catalog
Three ways to buy.
Start with clean voice. Add labels for duplex behavior. Add context lanes only when the model needs them.
Voice Core
Separated multi-speaker audio for ASR, diarization, and speaker separation.
- Isolated speaker channels
- Mixed reference audio
- Transcript + timing
- Session manifest
Duplex Plus
Voice Core with labels for real-time conversation behavior.
- Overlap windows
- Interruptions + barge-ins
- Tone + emotion tags
- Speech acts
Custom Context
Add consented context lanes around the same voice session.
- Screen video
- Keyboard + mouse input
- Facecam / affect
- Game events
Lane
Voice Core
Duplex Plus
Custom Context
Channel-separated audio
Included
Included
Included
Transcript + diarization
Included
Included
Included
Overlap / interruption labels
Optional
Included
Included
Tone / emotion tags
Optional
Included
Included
Screen video
-
-
Add-on
Keyboard + mouse input
-
-
Add-on
Facecam / affect
-
-
Add-on
Game state / events
-
-
Add-on
Governance
Consent and rights are part of the voice product.
- Participant consent covers voice recording, data processing, licensing scope, and model-use permissions.
- Each delivery includes consent-scope metadata, redaction notes, participant handling rules, and retention terms.
- Collections can exclude private messages, account credentials, desktop exposure, or any non-consented modality.
- Face, screen, and input lanes are opt-in add-ons, not required for the core voice dataset.
Access Process
How voice buyers get data.
-
01
Scope the voice corpus
Define speakers, languages, sample rate, labels, environments, consent terms, and delivery format.
-
02
Review sample and data card
Inspect separated tracks, mixed reference audio, transcript labels, manifest, and QA assumptions.
-
03
License or commission
License an existing voice package or commission a custom capture protocol for your target behavior.
-
04
Receive delivery
Get secure transfer, audio files, manifests, transcripts, QA report, and loader/schema support.
Contact Us
Tell us what voice data your lab needs.
Send the model target, speaker count, language scope, sample-rate requirements, label needs, consent constraints, and whether you need context lanes like screen, facecam, keyboard, or game events.