Domain Adaptation for AI Voice Agents (Vocabulary, ASR, TTS) in 2026

TL;DR — A general ASR mispronounces "metformin" 12% of the time and "amortization" 18% of the time. Domain adaptation — vocabulary biasing on the ASR + a pronunciation lexicon on the TTS + (optionally) acoustic LoRA — drops Word Error Rate 2–30 points in regulated verticals. It's the difference between a real voice agent and a science demo.

What it does

Three layers need adaptation:

1. ASR vocabulary biasing — at recognition time, boost the prior on a list of domain terms (drug names, plan names, internal codes).
2. TTS pronunciation lexicon — explicit <phoneme> overrides for brand names and rare words.
3. Acoustic LoRA (optional) — fine-tune the ASR or TTS encoder on domain audio if the dialect or recording channel is unusual (call-center mu-law audio, accented English, etc).

How it works

flowchart TD
  AUDIO[Caller audio] --> ASR[ASR with biasing]
  VOCAB[Domain phrase list] --> ASR
  ASR --> TEXT[Transcript]
  TEXT --> LLM[Agent LLM]
  LLM --> TTS_TEXT[Agent reply]
  TTS_TEXT --> TTS[TTS with lexicon]
  LEX[Pronunciation lexicon] --> TTS
  TTS --> SPEECH[Audio out]

CallSphere implementation

CallSphere runs 6 verticals · 37 agents with hard latency budgets (<800 ms TTFT). Domain adaptation we run today:

• Healthcare — vocabulary biasing on a 1,200-term list (drug names, ICD-10 phrases, plan names). WER on drug names down from 14% → 3%. Post-call analytics still on GPT-4o-mini but ASR feeds it cleaner text.
• Behavioral Health — TTS lexicon for crisis hotline names ("988", "SAMHSA") + de-escalation phrasing aliases. Latency-sensitive: we use vocabulary biasing only, no acoustic LoRA.
• Salon / Dental — TTS lexicon for hyphenated brand names ("Olaplex", "Invisalign") and stylist first names per location.
• OneRoof real-estate (OpenAI Agents SDK) — vocabulary biasing on neighborhood + street name lists per market.
• Behavioral Health PHI redactor — open-source LoRA on Whisper-large-v3 for telephony-band audio (8 kHz, mu-law).

Across 90+ tools · 115+ DB tables, the vocabulary list is generated from the database at deploy time — 90+ tools mean 90+ schema names that need to be pronounceable. Plans: $149 / $499 / $1,499, 7-day free pilot, 22% affiliate.

Build steps with code

# Deepgram domain biasing
deepgram.transcribe(
    audio,
    model="nova-3",
    keywords=["metformin:5","keytruda:5","prior-auth:3"],   # weight per term
    language="en-US",
)

# OpenAI Realtime + custom lexicon
session.update({
  "type":"session.update",
  "session":{
    "input_audio_transcription":{"model":"whisper-1"},
    "instructions":"Pronounce 'Olaplex' as oh-LAP-lex.",
  }
})

# Acoustic LoRA fine-tune of Whisper for telephony audio
from peft import LoraConfig, get_peft_model
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3")
peft = LoraConfig(r=8, lora_alpha=16, target_modules=["q_proj","v_proj"], lora_dropout=0.05)
model = get_peft_model(model, peft)
# train on (mu-law-resampled-audio, transcript) pairs

Pitfalls

• Over-biasing — too-high keyword weights cause the ASR to hallucinate domain terms in clean speech.
• Lexicon drift — TTS lexicons rot; refresh quarterly or whenever a new product launches.
• Forgetting per-region lists — neighborhood names, agent first names, and provider names are local. Generate per-tenant.
• LoRA on the wrong layer — Whisper LoRA on attention QV is the safe default; targeting feed-forward breaks transcription.
• Skipping noise-augmented training — call-center audio is noisy; train with reverb + telephony codec augmentation or your LoRA over-fits clean studio audio.

FAQ

Q: How big a vocabulary list is too big? Most ASRs handle 500–5,000 keywords cleanly. Past 5,000 you start eroding general performance.

Q: Vocabulary biasing vs acoustic LoRA? Biasing for terms; LoRA for accent/channel/dialect. They're orthogonal — combine them.

Q: Does this work with Whisper? Whisper supports prompt-based biasing (the initial_prompt parameter); acoustic LoRA needs HuggingFace + PEFT.

Q: How do I measure WER per vertical? Hold out 200 transcribed-and-corrected calls per vertical and compute WER weekly. Track per-term error rate too.

Q: When to skip acoustic LoRA? If your audio channel is clean and your accent distribution matches the base model's training data, biasing alone is enough.

Sources

• Deepgram — What Is a Voice AI Agent: 2026 Guide
• Vellum — Top 10 AI Voice Agent Platforms Guide 2026
• Inworld — Best Voice AI for Enterprise 2026
• NextLevel.ai — Best Speech-to-Text Models 2026
• OpenAI — Next-Generation Audio Models