CallMissed logoCallMissed
Back to all models
Speech to Textvoice-agentsstreamingturn-detection

Ink 2

by Cartesia · Released Undisclosed

Cartesia's top-ranked speech-to-text model for voice agents — 8% WER on AppTek's 14-accent call-centre benchmark, ahead of Deepgram Flux and ElevenLabs. Self-detects turns. Streaming only (voice sessions, not file transcription) and English only. $0.54 per audio hour.

Speech to Text

Ink 2

Powered by Cartesia · Streaming ASR with model-native turn detection (Cartesia Ink)

Context Window

N/A

Parameters

Undisclosed

Max Output

N/A

Category

Speech to Text

Overview

Ink 2 is Cartesia's best speech-to-text model for voice agents, and it is built for one job: live conversation. On AppTek's 14-accent call-centre benchmark it records 8% word error rate, ahead of Deepgram Flux at 10% and ElevenLabs at 12% — a meaningful margin on exactly the accented, noisy telephony audio that breaks weaker recognizers.

It detects turns itself. Rather than bolting a separate voice-activity or turn detector on top of the transcript, Ink 2 decides when a speaker has finished, which removes a layer of latency and tuning from the agent pipeline and produces more natural interruption handling.

Two constraints define where Ink 2 fits, and both are hard limits rather than preferences. First, it is streaming only: it runs on live voice sessions and is not accepted by the file-transcription endpoint, so it cannot be used to process recorded audio. Second, it is English only. Sending non-English audio does not return an error — it returns confidently wrong English text, which is worse than a failure, so do not select it for a non-English deployment.

Use Ink 2 when you are building an English-language voice agent and transcript quality on real telephony audio is the thing that matters most: contact-center agents, English call-centre automation, and turn-based conversational products. For multilingual work, for any Indic language, or for transcribing files, use ink-whisper instead — it covers 100 languages and works on both surfaces at a third of the price.

Pricing

MetricPrice
Price /hour₹54.0000

1 credit = ₹1 = $0.01 USD. Transparent per-call pricing — you pay only for what you call, with no seat fees or minimums.

Key Highlights

  • Cartesia's #1-ranked STT for voice agents
  • 8% WER on AppTek's 14-accent call-centre benchmark
  • Model-native turn detection — no separate turn detector
  • Streaming only, English only

Benchmarks

BenchmarkScore
AppTek 14-accent WER8%
LanguagesEnglish
Price$0.54
DeploymentStreaming only

Technical Details

  • English only — non-English audio is not supported
  • Streaming only — the file-transcription endpoint does not accept this model
  • Model-native turn detection (no separate turn detector required)
  • 8% WER on AppTek's 14-accent call-centre benchmark
  • Selected per voice session via stt_model

Strengths

  • Best-in-class WER on accented call-centre audio
  • Beats Deepgram Flux and ElevenLabs on the AppTek 14-accent benchmark
  • Self-detects turns, removing a pipeline layer and its latency
  • Purpose-built for live voice agents

Limitations

  • Streaming only — not available for file transcription (use ink-whisper for recorded audio)
  • English only — non-English audio returns incorrect English text rather than an error, so do not use it for multilingual or Indic deployments
  • Paid plans only (not in the free tier)
  • Three times the price of ink-whisper at $0.54 per audio hour

Use Cases

English voice agentsContact-center automationTurn-based conversational productsLive English call transcription

API Example

curl https://api.callmissed.com/v1/voice/sessions \
  -H "Authorization: Bearer cm_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "stt_model": "ink-2",
    "language": "en"
  }'

Endpoint: POST /v1/voice/sessions (streaming only) · Model ID: ink-2

Try Ink 2 now

Get 1000 free API credits on signup. No credit card required.

Start free Book a DemoRead docs