Skip to main content

CartesiaAudio modelActive

Cartesia Ink-2

Popularityno data yet

n/ano search or usage data

Summary

Cartesia Ink-2 is a streaming ASR model for voice agents. Cartesia reports about 0.1 seconds from the end of speech to the final transcript (TTFT), not to the first partial result. It adds semantic endpointing and structured entities to transcription.

The ink-2 endpoint is available through the Cartesia API and Play. Its launch named English, while a newer product page also lists Spanish, French, Hindi, and Japanese. The official sources do not give a consistent language list. Billing uses three credits per second of audio, with the dollar price depending on the plan.

GraduallyBench Audio

Cartesia Ink-2

Partially tested
  • The test coverage is incomplete.
  • The route does not confirm the exact model version.
  • Silence was not reliably evaluated.
  • The number of actual provider requests is not attested.
  • Only the indicated language subset was tested.
Show test details

Speech recognition

Public demo · English

15 valid outputs from 20 attempted and 20 planned test slots.

Word error rate (WER)
2.68 %(12 / 447 reference words)English
Character error rate (CER)
1.02 %(24 / 2,350 reference characters)English
Silence assessment
Not availableEnglish

Technical measurement

Public demo · English

1 valid output from 1 attempted and 1 planned test slots.

  • Only a short functional test is available.
  • The functional test does not assess content quality.
  • Only a technical test is available.
  • The selection of test inputs is limited.

No further scored metric is available for this test.

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 184 published benchmark results

Page 1 of 16

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

VoiceCodeBench: CTEM (Show measurement, test conditions, and source)
Source value
89.541
Metric
canonical token/entity match
Unit
%
Benchmark version
1
Category
general
Task
ctem
Task label
CTEM
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Vals AI
Status
active
Retrieved at
2026-11-07
Methodology
Source model ID: cartesia/ink-2. Source model name: Cartesia Ink-2. Provider reported by source: Cartesia. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-09-24. Latency in seconds: 69.197. Standard error: 0. Source ID: vals-ai. Evidence capture ID: vals-voice-code-bench-2026-11-07-7753d162bfe3. Source payload hash: 7753d162bfe3fa3470d586da3917bf8ddd96f14e10f9aa8f03dd517b8d5f003b. Parser version: vals-astro-v4. Imported at: 2026-11-07. Source cohort: vals-voice-code-bench:1:ctem
Source Vals AI
VoiceCodeBench: Overall (Show measurement, test conditions, and source)
Source value
62
Metric
task success rate
Unit
%
Benchmark version
1
Category
general
Task
overall
Task label
Overall
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Vals AI
Status
active
Retrieved at
2026-11-07
Methodology
Source model ID: cartesia/ink-2. Source model name: Cartesia Ink-2. Provider reported by source: Cartesia. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-09-24. Latency in seconds: 69.197. Standard error: 0. Source ID: vals-ai. Evidence capture ID: vals-voice-code-bench-2026-11-07-7753d162bfe3. Source payload hash: 7753d162bfe3fa3470d586da3917bf8ddd96f14e10f9aa8f03dd517b8d5f003b. Parser version: vals-astro-v4. Imported at: 2026-11-07. Source cohort: vals-voice-code-bench:1:overall
Source Vals AI
Endpointing precision (Show measurement, test conditions, and source)
Source value
89
Metric
precision
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Cartesia Ink-2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Internal July 2026 endpointing benchmark
Context
Cartesia's reported values and dated Vals observations use distinct metrics and cohorts; no cross-source rank is inferred.
Source Cartesia Ink-2 announcement
Endpointing recall (Show measurement, test conditions, and source)
Source value
97
Metric
recall
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Cartesia Ink-2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Internal July 2026 endpointing benchmark
Context
Cartesia's reported values and dated Vals observations use distinct metrics and cohorts; no cross-source rank is inferred.
Source Cartesia Ink-2 announcement
Endpointing F1 (Show measurement, test conditions, and source)
Source value
93
Metric
F1
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Cartesia Ink-2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Internal July 2026 endpointing benchmark
Context
Cartesia's reported values and dated Vals observations use distinct metrics and cohorts; no cross-source rank is inferred.
Source Cartesia Ink-2 announcement
Coval pooled corpus-level STT word error rate (Show measurement, test conditions, and source)
Source value
3.5860138094
Metric
Pooled corpus-level speech-recognition word error rate, Coval pooled_value
Unit
%
Benchmark version
Coval rolling 30-day window, re-measured daily; route last measured 2026-10-16
Category
audio
Direction
Lower is better
Sample
26673
Source type
official-leaderboard
Evaluator
Coval
Status
superseded
Retrieved at
2026-10-16
Methodology
Official benchmark leaderboard. Coval endpoint key cartesia/ink-2, exact API model ID ink-2, hosted by Cartesia. The route identifies Ink 2 separately from Cartesia Sonic speech-synthesis models. Rolling 30-day window, re-measured daily.
Context
This is Coval's pooled corpus-level STT WER, not the separate per-clip mean_value or a rounded leaderboard headline. Pooled WER and mean per-clip WER are different aggregations and should not be compared as the same score. The WER row's sample_count is used, not another metric's count or a roster total. The result describes Coval's hosted 30-day endpoint cohort, not local or native performance.
Source Coval
Cekura Converse-STT, Cartesia Ink 2 Pipecat WER (Show measurement, test conditions, and source)
Source value
2.39
Metric
word error rate (WER) on Pipecat public short-clip speech-to-text set
Unit
%
Benchmark version
Cekura Converse-STT page last updated 2026-09-16, retrieved 2026-10-17
Category
audio
Direction
Lower is better
Sample
1000
Lower confidence bound
2.13
Upper confidence bound
2.65
Source type
independent-evaluator
Evaluator
Cekura Bench
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Streaming STT model row named Cartesia Ink 2. Pipecat dataset contains 1,000 public clips; the row reports 1,000 usable model outputs. Cekura labels the uncertainty estimate a likely 95% range. The served endpoint build is not pinned.
Context
The published interval is Cekura's likely 95% range, not a separately documented model confidence interval. The page states that per-model counts are usable outputs. This distinct public-clip WER observation is not the existing Coval, Vals VoiceCodeBench, or Cartesia vendor-reported WER.
Source Cekura Bench
Cekura Converse-STT, Cartesia Ink 2 Ocular WER (Show measurement, test conditions, and source)
Source value
3.09
Metric
word error rate (WER) on Ocular licensed real-phone-conversation speech-to-text set
Unit
%
Benchmark version
Cekura Converse-STT page last updated 2026-09-16, retrieved 2026-10-17
Category
audio
Direction
Lower is better
Sample
8
Lower confidence bound
2.43
Upper confidence bound
3.84
Source type
independent-evaluator
Evaluator
Cekura Bench
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Streaming STT model row named Cartesia Ink 2. Ocular provides eight licensed recordings from four real phone conversations; the row reports usable outputs for all eight. Cekura labels the uncertainty estimate a likely 95% range. The served endpoint build is not pinned.
Context
The published interval is Cekura's likely 95% range, not a separately documented model confidence interval. The page states that per-model counts are usable outputs. This distinct real-call WER observation is not the existing Coval, Vals VoiceCodeBench, or Cartesia vendor-reported WER.
Source Cekura Bench
Speko FLEURS clean read-English batch WER, cartesia:ink-2 (Show measurement, test conditions, and source)
Source value
11
Metric
Word error rate, mean per clip; lower is better
Unit
%
Benchmark version
Speko FLEURS read-English batch cohort; No measurement date is stated for this exact model row. The board-level measured date 2026-07-03 is not assigned to this route.
Category
speech-to-text
Direction
Lower is better
Sample
50
Source type
independent-evaluator
Evaluator
Speko independent speech benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. English read speech, FLEURS clean test, batch API, exact Speko route row. The Speko board’s global batch cohort is n=50; exact model-specific backend revision and API controls are not disclosed. Its board-level measured date 2026-07-03 is not assigned to this model row.
Context
This is an independent 50-clip FLEURS batch result, not the same streamed or conversation dataset as the other rows. Exact provider backend/revision is not pinned. A new retrieval did not rerun the measurement; the exact row date is stated separately only if published.
Source Speko independent speech benchmark
Speko streaming FLEURS test+dev+train English WER, cartesia:ink-2 (Show measurement, test conditions, and source)
Source value
10.12
Metric
Word error rate on streaming transcript; lower is better
Unit
%
Benchmark version
Speko STT streaming FLEURS test+dev+train run; 2026-10-01.
Category
speech-to-text
Direction
Lower is better
Sample
1541
Source type
independent-evaluator
Evaluator
Speko independent speech benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. English read FLEURS clips streamed at real-time pace through the exact Speko provider route; source combines FLEURS test, development and training clips. Speko reports n=1,541 clips and measurement date 2026-10-01.
Context
The source explicitly combines FLEURS test, development and training clips, so this is not a clean held-out test set. It does not publish a confidence interval or immutable provider build; its displayed cell text supplies the exact route and evaluated corpus.
Source Speko independent speech benchmark
Speko EdAcc spontaneous accented English conversation WER, cartesia:ink-2 (Show measurement, test conditions, and source)
Source value
16.6781
Metric
Word error rate over spontaneous accented conversational speech; lower is better
Unit
%
Benchmark version
Speko EdAcc spontaneous-conversation realtime run; 2026-10-01.
Category
speech-to-text
Direction
Lower is better
Sample
283
Source type
independent-evaluator
Evaluator
Speko independent speech benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Spontaneous accented English conversation (EdAcc), 60 speakers, 283 clips, two real-time passes, exact Speko provider route. Speko reports n=283 clips and measurement date 2026-10-01; the two passes are not treated as independent examples.
Context
The Speko source displays a range (e.g. 14.4–19.9%) in the same cell but does not label its interval type or confidence level; it is not imported as a confidence interval. It reports 60 speakers and two passes; the N is 283 source clips, not 566 passes. No immutable backend build is published.
Source Speko independent speech benchmark
Coval Cartesia Ink 2 word error rate (WER), 30-day mean (Show measurement, test conditions, and source)
Source value
3.6
Metric
Word error rate (WER)
Unit
%
Benchmark version
Coval rolling 30-day fixed-public-audio results; exact route last run 2026-10-17 10:00 UTC
Category
audio
Direction
Lower is better
Source type
official-leaderboard
Evaluator
Coval
Status
superseded
Retrieved at
2026-10-19
Methodology
Official benchmark leaderboard. Coval exact model route Ink 2, hosted by Cartesia using its official API; daily evaluation of fixed public speech across clean, accented, noisy, reverberant, far-field, clipped, and phone-codec conditions; rolling 30-day results; exact-model page says last run 2026-10-17 10:00 UTC.
Context
The page header states 133,365 samples over 30 days, while a separately opened Oct 19 Coval dashboard row for Ink 2 shows 26,667 samples together with conflicting metrics. The sources do not explain whether the counts share a denominator, so sampleCount is left unknown. That undated dashboard row was not used; the exact route-page run is retained with its stated Oct 17 timestamp. No confidence interval is shown.
Source Coval

Page 1 of 16

All specifications

Architecture

Model class
Streaming speech recognition
Source: Model class

Input and output

Input
Audio
Source: Input
Output
Transcript and structured entities
Source: Output
Languages
English at launch according to the announcement; the current Ink product page also lists Spanish, French, Hindi, and Japanese. Language-coverage claims are not consistent.
Source: Languages
Time to final transcript
Provider-reported 0.1 seconds from the end of speech to the final transcript (TTFT), not to the first partial result
Source: Time to final transcript

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Cartesia API and Play
Source: Access
Shutdown
None announced
no source

API

API model ID
ink-2
Source: API model ID

Further details

Features
Semantic endpointing and structured entity recognition
Source: Features

Prices

Price
3 credits per second of audio; the dollar price and overage charges depend on the planSource

Sources and data date

Show all 23 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

EvaluatorTest seriesComposite indicesRecorded active measurements
Vals AI103

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 20 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Unresolved

benchmark-cohorts

Note

Cartesia's reported WER and endpointing results, current Vals WAcc/CTEM/TSR rows, and the dated 2026-09-29 Vals WER/CTEM/TSR capture use distinct metrics or cohorts. They remain separate and are not overwritten or ranked together.

Unresolved

exact-serving-snapshot-and-unpublished-method-details

Note

The captured external source reports a benchmark-specific model or hosted route but does not establish an immutable currently served checkpoint. Unpublished row-level denominators and confidence intervals remain unknown; sample counts were entered only for cells that state them.

Additional sourceCartesia Ink-2 announcement (retrieved October 3, 2026; October 4, 2026; October 22, 2026) · Cartesia Ink-2 announcement · Editorial description reviewed October 3, 2026
Additional sourceCartesia Ink product page (retrieved October 3, 2026) · Cartesia Ink product page · Editorial description reviewed October 3, 2026
Additional sourceCartesia pricing (retrieved October 3, 2026) · Cartesia pricing · Editorial description reviewed October 3, 2026
Additional sourceVals AI (retrieved November 7, 2026)
Additional sourceCoval (retrieved October 16, 2026; October 17, 2026; October 19, 2026; October 20, 2026; October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 5, 2026; November 6, 2026; November 7, 2026)
Additional sourceCekura Bench (retrieved October 17, 2026)
Additional sourceSpeko independent speech benchmark (retrieved October 17, 2026)
Additional sourceCekura Bench (retrieved October 23, 2026)
Additional sourceModel metadata (retrieved October 16, 2026)12
Additional sourceModel metadata (retrieved October 17, 2026)12
Additional sourceModel metadata (retrieved October 17, 2026; October 28, 2026)
Additional sourceModel metadata (retrieved October 28, 2026)1234
Additional sourceModel metadata (retrieved November 5, 2026)123
Additional sourceModel metadata (retrieved November 6, 2026; November 7, 2026)123