Cartesia is best known for Sonic, its text-to-speech model, but it also sells streaming speech-to-text under the Ink name. The current model, Ink-2, is tuned for voice agents: it detects when a caller starts and stops talking, handles phone audio and background noise, and returns final text fast. It transcribes. It does not translate, and its docs describe no speaker labels.
The LecSync Realtime Speech API is built for a different job: live captions and meeting transcripts. One WebSocket returns the transcript, a translation between any two of 60 languages, speaker labels and running minutes, at $0.006 per audio minute ($0.36 per hour).
All Cartesia figures below come from its official pages, read on October 19, 2026: the pricing page, the pricing docs, the Ink-2 model page, older STT models, STT endpoint comparison, concurrency limits and the Ink product page. Every price is in US dollars, so no currency conversion is involved.
The short answer
For a voice agent in English, Spanish, French, Hindi or Japanese, Ink-2 is a strong pick. Its built-in turn detection replaces a separate voice activity model, and it accepts 8 kHz μ-law phone audio directly. It costs 3 credits per second, which works out to roughly $0.40 to $0.70 per hour depending on the plan.
For captions that need a second language, speaker names, or more than five languages on a fast model, LecSync is the simpler choice. At 1,000 hours a month, Ink-2 transcription alone costs about $405 on Cartesia's Scale plan. LecSync costs $360 with translation, speakers and minutes included.
How each option works
LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, the language pair and the features you want. It gets back a stream URL and a one-time token, and the client then streams 16 kHz mono PCM. On that one connection you receive preview and final transcript segments, translations paired by segment ID, speaker labels with millisecond timestamps (speakers: true) and sections of minutes as the meeting goes (digest: true). The first word reaches the screen about 0.28 seconds after someone starts speaking, and translation adds roughly 88 ms. If the network stalls, backlog is capped at 2 seconds.
Cartesia Ink-2, auto mode. wss://api.cartesia.ai/stt/turns/websocket runs Ink-2 with turn detection built in. The server sends turn.start, turn.update, turn.eager_end, turn.resume and turn.end events, each carrying the transcript for the current turn. Text is final when emitted; the model does not revise earlier output. Ink-2 covers English, French, Hindi, Japanese and Spanish and detects which one is being spoken. Cartesia quotes a time to final transcript of 0.1 seconds.
Cartesia, manual mode. /stt/websocket leaves turn handling to you: your own voice activity detection sends finalize when the user stops speaking, and the socket returns transcript deltas with word-level timestamps. This endpoint also runs the older ink-whisper model, which lists about 100 languages and costs a third as much as Ink-2.
Both Cartesia endpoints accept raw PCM in six encodings, from 16-bit integer to 32-bit float plus μ-law and A-law, at 8 to 48 kHz. Up to 100 key terms (1,200 characters in total) can be passed when the socket opens, but not changed mid-stream.
Side-by-side facts (checked October 19, 2026)
| LecSync Realtime Speech API | Cartesia Ink-2 | Cartesia ink-whisper | |
|---|---|---|---|
| Built for | Live captions, meetings, lectures | Voice agents | Lower-cost transcription |
| Price | $0.006/min ($0.36/hr), all features | 3 credits/sec (≈ $0.40–$0.70/hr by plan) | 1 credit/sec (≈ $0.13–$0.23/hr by plan) |
| Languages | 60, auto-detected | 5: en, es, fr, hi, ja | About 100 |
| Live translation | Any two of 60 languages, same socket | Not offered | Not offered |
| Speaker labels | Every final segment | Not documented | Not documented |
| Turn detection | Not offered (caption-oriented) | Built in, five event types | Manual finalize |
| Live minutes | digest: true | No | No |
| Audio input | 16 kHz mono PCM16 | PCM, μ-law, A-law, 8–48 kHz | Same |
| Concurrent streams | Not published | 8 / 12 / 20 / 60 (Free / Pro / Startup / Scale) | Same pool |
| Regions | 6 gateway regions | Auto-routed; dedicated US, EU, UK, India, Australia on Enterprise | Same |
| Silence | Accepted silence billed; empty connections free | Billed | Billed |
Cartesia closes STT sockets that sit idle for 3 minutes, and an idle socket still counts against your concurrency limit.
What the credits actually cost
Cartesia sells a monthly credit pool. Ink-2 uses 3 credits per second, or 10,800 credits per hour of audio. Credits are shared with text-to-speech, so any TTS you run draws from the same pool.
| Plan | Monthly price | Credits | Ink-2 hours included | Overage per 1M credits | Ink-2 overage per hour |
|---|---|---|---|---|---|
| Free | $0 | 20,000 | ~1 h 51 min | Not available | — |
| Pro | $5 | 100,000 | ~9 h 16 min | $65 | $0.70 |
| Startup | $49 | 1,250,000 | ~115 h 44 min | $45 | $0.49 |
| Scale | $299 | 8,000,000 | ~740 h 44 min | $38 | $0.41 |
Dividing the Scale price by its included hours gives about $0.40 per Ink-2 hour. Cartesia's FAQ rounds this to "$0.39 on the Scale plan." Overage has to be switched on, or requests fail once credits run out. Unused credits roll over up to twice the monthly allotment. Enterprise pricing is custom.
What 100 and 1,000 hours cost
| Monthly audio | LecSync | Cartesia Ink-2 (cheapest plan) | Cartesia ink-whisper (cheapest plan) |
|---|---|---|---|
| 100 hours, same-language transcript | $36 | $49 (Startup) | $21.90 (Pro + overage) |
| 1,000 hours, same-language transcript | $360 | $405.40 (Scale + overage) | $154.75 (Startup + overage) |
| 1,000 hours, English speech with translated captions | $360 | ≈ $1,475 with Google Cloud Translation (estimated) | ≈ $1,225 with Google Cloud Translation (estimated) |
| Speaker labels or live minutes | Included | Need another provider | Need another provider |
The translation row assumes you send only finished turns to Google Cloud Translation's default model at $20 per million characters after the first 500,000 free each month. English speech at about 150 words per minute is roughly 54,000 characters an hour, so 1,000 hours is about $1,070 in translation. That volume is an estimate. For the same math across more vendors, see what 1,000 hours cost once translation, speakers and summaries are added.
When Cartesia is the better pick
- Voice agents. Native turn events with an eager-end signal let your LLM start before the caller fully finishes. LecSync has no turn-taking events.
- Phone audio. μ-law and A-law at 8 kHz go straight in. LecSync v1 accepts only 16 kHz PCM16, so telephony audio needs resampling first.
- Cheap same-language transcription. On ink-whisper, 1,000 hours costs about $155 on the Startup plan, well under LecSync's $360.
- Compliance and deployment. Cartesia lists SOC 2 Type II, HIPAA, GDPR and PCI, and offers on-prem, VPC and zero data retention under Enterprise contracts.
- One vendor for speech in and speech out. If you already use Sonic for TTS, the same credits cover Ink.
When LecSync fits better
- Translated captions. Cartesia has no translation on any STT endpoint. LecSync translates between any two of its 60 languages on the transcription stream, and picks the direction per segment, so a bilingual room works without switching settings. Why this saves work is covered in built-in translation vs a DIY STT + MT pipeline.
- Knowing who said what. LecSync labels every final segment with a speaker and audio-aligned timestamps. See building a meeting client with speakers and timestamps.
- Languages beyond five on a current model. Ink-2 covers five. LecSync covers Chinese, Korean, German, Arabic, Russian and the rest of its 60 languages with the same features.
- Minutes while the meeting runs.
digest: truepushes confirmed sections over the same socket. - Long sessions over shaky networks. The 2-second backlog cap keeps captions near live after a stall, as described in how LecSync caps backlog at 2 seconds.
- No plan to size. No monthly fee, no credit pool, no overage switch. You pay for accepted audio, down to fractions of a second.
FAQ
Does Cartesia Ink support real-time translation?
No. As of October 19, 2026, Cartesia's speech-to-text endpoints return transcripts only, and its STT docs describe no translation option. To show translated captions you would send each finished turn to a separate translation API. LecSync returns translations on the same WebSocket as the transcript, at the same $0.006 per minute.
How much does Cartesia Ink-2 cost per hour?
Ink-2 uses 3 credits per second of audio, or 10,800 credits an hour. On the $299 Scale plan that is about $0.40 per hour (Cartesia's FAQ says $0.39), about $0.42 on Startup and $0.54 on Pro. Overage runs $0.41 to $0.70 per hour. Silence is billed too.
Which languages does Cartesia Ink-2 support?
English, Spanish, French, Hindi and Japanese, detected automatically. The older ink-whisper model lists about 100 languages but runs only on the manual-finalize endpoint, without built-in turn detection. LecSync transcribes 60 languages and translates between any two of them.
Can Cartesia Ink label speakers?
Cartesia's STT documentation and API reference describe no speaker diarization for Ink-2 or ink-whisper as of October 19, 2026. Responses carry the transcript and, on the manual endpoint, word timestamps. LecSync adds a speaker label and timestamps to every final segment when speakers: true is set.
Bottom line
Choose Cartesia Ink-2 when you are building a voice agent in one of its five languages, especially over phone lines, or ink-whisper when you need the cheapest same-language transcript and can handle turn detection yourself. Choose LecSync when the output is captions people read: translated, labelled by speaker, with minutes building up during the session, for one flat $0.006 per minute. You can open a test stream from the developers page with a single POST /v1/realtime/connect.