Voice AI and phone agents
Voice is metered per minute and the cost accumulates per component. What decides the choice is whether the per-minute price is consolidated, whether the Japanese is good enough to use, and whether the latency reads as conversation.
Published · updated
How we ranked these
Stated before the list, so you can disagree with the weighting rather than the result.
The ranking
- 01
Deepgram4.5従量課金
Speech-to-text and text-to-speech APIs from $0.0043 a minute, with a self-hosted option on enterprise.
The cost leader for bulk pre-recorded transcription. Test Japanese accuracy with the free credit.
- 02
AssemblyAI4.5従量課金
Transcription and speech understanding from $0.15 an hour of pre-recorded audio.
Strong where redaction and understanding matter. Model streaming cost on real session lengths.
- 03
Cartesia4.4$5/mo
Low-latency speech synthesis and recognition, plus managed voice agents.
Worth a direct latency comparison. The entry price makes the test trivial.
- 04
ElevenLabs4.7From $5/mo
The industry standard for synthetic speech: multilingual narration, voice cloning and dubbing.
The first place to look for professional narration, dubbing or voice cloning. Managing the consent and rights side is on you.
- 05
CoeFont4.1$350/mo
Japanese speech synthesis plus real-time AI interpretation that follows the speaker's context.
Strong for Japanese voiceover and for keeping mixed-language meetings moving. Not for high-stakes interpretation.
- 06
Bland AI4.2$0.14/分
Voice AI for phone calls that bundles the model, transcription and speech into a single per-minute rate.
The easiest voice platform to budget for. Test Japanese audio before committing.
- 07
Vapi4.2従量課金
Infrastructure for voice AI agents that answer the phone, billed by the minute.
The right foundation for a phone-answering agent. Budget the pilot for tuning latency, and settle the consent question first.
- 08
Synthflow4.1Contact sales
Enterprise voice agents sold as an annual contract that includes telephony, integrations and security review.
Enterprise-only by design. Use Bland or Vapi to pilot, then consider this at scale.
- 09
Voiceflow4.2Contact sales
Design and operate voice and chat agents, with your choice of model provider or your own.
A sensible choice for agencies and teams that treat agents as production software.
- 10
Hume AI4.1Contact sales
Evaluation infrastructure for voice AI, scoring timing, tone, interruption and accent against a published benchmark.
Relevant if you are shipping voice at scale. Premature before that.
Questions we get
- What does a thousand minutes cost?
- It depends on the construction. Bland consolidates to a single per-minute rate, with Start at $0.14 a minute, so roughly $140. Vapi bills per component and its own calculator produces $82 to $129. Telephony is separate in both cases.
- How do we verify Japanese accuracy?
- Only with your own recordings. Published accuracy figures use incomparable conditions. Deepgram offers $200 and AssemblyAI $50 in credit, so run the same audio through both.
- How much can be delegated?
- Routine questions like opening hours and location, taking bookings, and first-line triage. Keep initial complaint handling and anything that fixes an amount with a person.