Anyreach
Research lab

The science behind AI that speaks.

Voice AI is harder than it looks. Voicemail detection and sub-second latency make or break a real conversation. We build the stack ourselves, and publish it.

0Research studies
0Solved in production
0+Benchmarks run
Interspeech 2026 · Peer-reviewedDualTurn: learning turn-taking from dual-channel generative speech pretrainingRead the paper →
The call lifecycle

Every call is five problems in a row.

Pick a stage to see the research behind it.

Research library

Built, benchmarked, in production.

Core voice pipelineanyreach-asr

Speech recognition for voice agents

5.26%word error rate (batch)Core voice pipelineAnyreachTTS

Human-like speech for voice agents

3BOrpheus LLaMA 3.2 backboneCore voice pipelineAnyLingual

Low-latency speech translation

0.76 slatency (Small)Core voice pipelineTurn-taking

Three open models for conversation dynamics

+53%turn prediction vs best prior modelCore voice pipelineVoicemail detection

When your brand speaks, make sure it lands

94.8%live delivery (from 83.2%)Quality & evaluationAVM

A unified quality score for voice agents

≥0.8target human-QA correlationQuality & evaluationQuality & evaluation

Testing, simulation and assessment at scale

100%of calls assessedPaper · Interspeech 2026DualTurn

Learning turn-taking from dual-channel generative speech pretraining

0.633wF1 on Switchboard (VAP 0.389)
Paper · Interspeech 2026

DualTurn · Learning turn-taking from dual-channel generative speech pretraining

Shangeth Rajaa (Anyreach AI). Accepted at Interspeech 2026. arXiv:2603.08216.

0.000wF1 on Switchboard (VAP 0.389)
0.000backchannel F1 (VAP 0.000)
~0 mson a single CPU
0 msearly on turn boundaries

Agent-action wF1, Switchboard (higher is better)

DualTurn (0.5B)0.633
VAP (5.8M)0.389

What we built

A 0.5B backbone pretrained to predict both speakers’ future audio, then lightweight heads for five agent actions.

Why it matters

The first use of speech-to-speech generative pretraining for explicit turn-taking in modular ASR-LLM-TTS pipelines.

  • “The backbone is the vessel, not the source.”
  • 56% of turn-end signal is semantic, 26% prosody
  • Continuous embeddings beat discrete tokens
  • Code, model and datasets are open
Read the full study →Read on arXiv →Code →Model →
The problem map

Eight hard problems. Four solved, four next.

0/8solved in productionMaking sure messages land94.8% deliveryKnowing who answeredComing soonUnderstanding every voice5.26% WERInstant responses~400 ms achievedReading customer intentComing soonNatural-sounding AIOrpheus LLaMA 3.2Staying compliantComing soonQuality at scale100% coverage
Making sure messages landVoicemail detection94.8% delivery
Knowing who answeredAnswer machine detectionComing soon
Understanding every voiceSpeech recognition5.26% WER
Instant responsesConversational latency~400 ms achieved
Reading customer intentIntent classificationComing soon
Natural-sounding AIVoice synthesisOrpheus LLaMA 3.2
Staying compliantCompliance & consentComing soon
Quality at scaleCall analytics & QA100% coverage
Want to see how we solve these?The research runs in every program we ship.
Talk to SalesExplore products