Voice agent buyer guide
Native Speech-to-Speech vs. Cascaded Voice Agents
Understand what changes when audio is handled directly by a realtime model rather than passing through separate transcription, language-model, and synthesis stages.
Prepare repeatable speech tests
Check your SSML and compare common provider differences with synthetic callback, price and pronunciation fixtures.
Open the free local SSML editor →What the terms mean
A cascaded pipeline exposes replaceable STT, LLM, and TTS components, which can improve provider choice and debugging but adds boundaries where timing and meaning can be lost. Native audio models can respond directly to audio and may preserve more prosodic context, but model behavior and cost can be less componentized.
- Evidence
- A primary document, observed pilot result, or contract term that supports a claim.
- Production path
- The complete route from caller and carrier through models, tools, transfer, logs, and human follow-up.
Decision record
Compare each option across the same operational criteria.
Swipe or scroll the table to see every column.
| Criterion | Native speech-to-speech | Cascaded or configurable | Evidence to collect |
|---|---|---|---|
| Buyer requirement | A feature claim | A pass/fail production behavior | Recorded scenario and artifact |
| Cost | Headline rate | Complete workload cost | Quote with every component |
| Risk | Policy statement | Tested control and owner | Configuration, log, and contract |
Reusable decision tool
Build the evidence pack
Check the work that must be complete before a decision.
Architecture changes the failure surface
A cascaded pipeline exposes replaceable STT, LLM, and TTS components, which can improve provider choice and debugging but adds boundaries where timing and meaning can be lost. Native audio models can respond directly to audio and may preserve more prosodic context, but model behavior and cost can be less componentized.
Architecture alone does not prove lower real-world latency, higher task completion, or accurate emotion recognition. Measure the full production path.
Test the behavior callers notice
Run interruptions, corrections, long pauses, accents, noise, numbers, names, and emotionally charged but safe scenarios. Record time-to-first-audio, overlap recovery, tool accuracy, transfer success, and reviewer preference.
Keep vendor claims separate from your measurements and document model/version changes.
Take this to every provider
Your evaluation checklist
- Diagram every candidate's audio path
- Record exact models and versions
- Measure interruption and recovery
- Measure tool completion and transfer
- Avoid treating emotion claims as ground truth
Use the decision tool above and the full pilot worksheet to record results against each shortlisted provider: Shortlist workspace.
Primary sources
Source summaries reviewed 2026-09-04.
OpenAI documents realtime audio input and output.
PolyAI documents ASR, language model, and TTS components.
Microsoft documents native GPT-Realtime and cascaded GPT-5-Chat voice modes.
Providers to investigate
See all 27 providers →These examples span the operating models covered in this guide.
Swipe or scroll the table to see every column.
| Providers | What to investigate | Considerations |
|---|---|---|
DocsBot Voice Agents
| Verify Native realtime audio model option with knowledge retrieval and tools on your real phone route; then test interruption handling and scheduled human transfer. | Voice Agents are in beta. Confirm model behavior, telephony, tools, and transfer rules on real calls. Profile sources · checked 2026-09-04 |
| Realtime API | Verify Native audio input/output with tool calling on your real phone route; then test Not publicly documented. | Not a turnkey contact center. Implement telephony operations, identity, tools, monitoring, evaluation, retention, and escalation. Profile sources · checked 2026-09-04 |
| Empathic Voice Interface | Verify Native speech-to-speech with prosody-sensitive processing on your real phone route; then test Not publicly documented. | Empathic positioning is not proof of accurate emotion inference. Test false interpretations, cultural variation, consent, sensitive uses, and escalation. Profile sources · checked 2026-09-04 |
| Ultravox | Verify Native speech-language model on your real phone route; then test Not publicly documented. | Open weights do not remove hosting, latency, safety, evaluation, voice, and telephony work. Review license and SLA. Profile sources · checked 2026-09-04 |
| Copilot Studio Voice | Verify Native GPT-Realtime mode or cascaded GPT-5-Chat mode on your real phone route; then test Not publicly documented. | Product names, licenses, previews, and architecture vary. Require a bill of materials for Copilot, Dynamics, Azure, telephony, and services. Profile sources · checked 2026-09-04 |
Before you decide
Questions buyers ask
Is native speech-to-speech always better?
No. It can remove fixed pipeline boundaries and preserve audio context, while a cascade can offer replaceable components and clearer debugging. The better choice is the one that passes your scenarios at acceptable cost and operational risk.
Can a voice model understand emotion?
Some vendors say their models use prosody or other vocal cues. That is not perfect emotion recognition. Test false interpretations, cultural and language variation, consent, and escalation in sensitive use cases.
Keep building your shortlist
Browse voice-agent providers →Compare candidates →Build an evidence-backed shortlist →All buying guides →Evaluate language support in real calls
Test recognition, speech quality, interruptions, language switching, tool use, and human transfer with your own calls in the required language. Availability can vary by model, plan, and phone route.
DocsBot documents multilingual voice workflows and Japanese support. Verify the exact voice, phone path, and transfer behavior in your pilot. Review DocsBot language information ↗
Other providers’ language claims reflect cited vendor material; confirm Japanese and other required languages directly with each provider.
Listing does not imply endorsement or a performance guarantee. This directory is published by DocsBot. Editorial policy