Cartesia Ink vs Giggy Speech to Text

Cartesia Ink 68

The service is closed and the SDKs are Apache-2.0. The Terms of Service cover the website, its subdomains and APIs, were last revised on 14 June 2024, describe a platform for synthesised sound and never mention transcription (12 of 30). The terms and privacy policy agree that content may be used for training with an opt-out. The DPA says Cartesia will not retain or use customer personal data beyond performing the services, which sits uneasily with the training licence. No retention period is given outside zero data retention, and the privacy policy says the services are designed for users in the United States only while regional endpoints are sold in four other regions (12 of 30). A deprecations page with dated sunsets and a stated promise to preserve fields for a pinned Cartesia-Version (15 of 20). The DPA promises 30 days' notice of sub-processor changes to subscribers and points to a list in the trust centre, which we could not read. Regional deployments are named for Enterprise only (10 of 20).

Giggy Speech to Text 30

Closed hosted service. The terms exist at /terms but draw by script and could not be read, so their clarity is unknown; the SDK, Pipecat and MCP repositories are ISC and the n8n node MIT (5 of 30). What could be read agrees with itself, no training on user data, an optional 12-hour retention flag, a seven-day window for Batch audio, and the EEA compute flag. The privacy policy, any DPA and a retention period for transcripts were not found or not readable (12 of 30). No deprecation policy or dated notices, only a statement that the legacy origin remains supported (3 of 20). No subprocessors or data locations are named; the EEA-adequate compute option implies processing outside the EEA by default without saying where (3 of 20). 5+12+3+3 = 23.

Cartesia Ink 73

Hosted reading. The status page at status.cartesia.ai (incident.io) lists Speech to Text as its own component in four regions (20). Its RSS feed, re-read on 1 November, now names each incident's components. In the 90 days from 3 August two incidents touch speech-to-text, both partial and read as minor, as this run reads one region or some requests. On 2 September some clients in the US regions saw connection errors that cleared on retry (Text to Speech and Speech to Text, US), and on 7 October customers in the US regions saw higher latency for some API requests, from 12:29 to about 18:23 (partial outage, Text to Speech, Speech to Text and Voice Cloning, US). The 3 September throttling named only Text to Speech in APAC, the 8 October errors only the Playground and Voice Agents, and the 15 October Playground outage says API services were unaffected. The 1 August API-wide incident has left the window. Minor only (20). Concurrency limits are published per plan, 8, 12, 20 and 60 concurrent STT requests, with a 3-minute idle timeout on sockets (15). Exceeding the limit returns 429 with the concurrency_limited code, and the SDKs retry 429 and 5xx twice with backoff on HTTP calls. No Retry-After header, backoff guidance or reconnection advice for a dropped socket was found (8 of 15). No SLA was found for any plan (0). ink-2 is marked Stable on the model page re-read on 1 November, with ink-preview as the preview line (10). 20 + 20 + 15 + 8 + 0 + 10 = 73.

Giggy Speech to Text 50

Hosted reading, as for deepgram-stt and cartesia-ink-stt. Giggy's own status page at https://giggy.ai/status lists one component, Text to speech, operational on 4 November 2026 with 98.61 per cent observed availability and 99.85 per cent completion over 14 days, refreshed every 30 minutes. It has no speech-to-text component and no incident log, so it shows nothing about this surface (5 of 20). No readable incident history, so the floor (5). Rate limits are published per credential, 60 Batch generation POSTs, 120 event-stream GETs, 600 reads and 120 MCP requests a minute, with eight live contexts per controller; Fast and Streaming POSTs have no per-credential limit and Batch queue capacity isn't given (15). 429 carries Retry-After with RateLimit-* headers, and the docs give a safe-retry rule for uploads, read the generation by UUID after a timeout or 504 and never upload the same recording again. REST transcription has no idempotency key, but the MCP transcribe_audio tool takes one (15). No SLA on the pricing page or in the docs; the terms could not be read (0). No beta or preview label on the API, the SDKs or the MCP server (10). 5+5+15+15+0+10 = 50.

Cartesia Ink 89

Model reading. The docs index links an OpenAPI file (latest.yml) and AsyncAPI files for the WebSocket endpoints, and each reference page embeds its part of them (25). llms.txt with a Markdown twin of every page (10). The comparison page says which of the three endpoints to use and when, and the turn detection guide explains each event and threshold (18 of 20). The AsyncAPI definition has enums for encoding and cartesia_version, minimum and maximum values on the four turn settings and required fields marked. model is a plain string on the auto endpoint (13 of 15). Example scripts for each endpoint in Python and TypeScript and an errors page with nine codes. model_sunsetted appears on the deprecations page and not in that table, and the docs say a wrong encoding or sample_rate might not return an error (11 of 15). Dated API versions through the Cartesia-Version header and a changelog grouped by month with no day dates (12 of 15).

Giggy Speech to Text 68

Model reading, with the API reference. The docs and llms.txt link an OpenAPI 3.1 file at https://api.giggy.ai/v1/openapi.json, but api.giggy.ai's robots.txt could not be read, so the file wasn't fetched. We credit only what the reference pages show of it, typed mode values, the privacy booleans and a typed error body, and there is no AsyncAPI file for the live socket (12 of 25). llms.txt and llms-full.txt (10). The docs say when to use each mode and transport, that MCP is not a live transport, that an OpenAI base-URL swap does not work for STT, and which options have no equivalent, more than most listings say about when not to use something (17 of 20). mode is an enum of three values and the input format is fixed, but language is a free code or name with no published list, and the MCP input schemas were read only from the repository's docs, not from the server (11 of 15). Examples in the SDK READMEs, the examples repository and a Postman collection, and an error table of ten statuses with type, code, message, status, request_id and param fields, though only one code value is named (11 of 15). Paths are under /v1 and the SDK and Pipecat repositories keep CHANGELOG.md files, dated in two of them, but there is no service changelog (7 of 15). 12+10+17+11+11+7 = 68.

Cartesia Ink 84

Read as a model, on the lines deepgram-stt, gladia-stt and cartesia-tts use for hosted speech models (recency 30, dated releases 10, deprecation notice 10, support 15, issues 10, SDKs 15, health 10). The cartesia-mcp server, which has a speech_to_text tool, released 0.28.1 on 30 October 2026 with a bug fix, after 0.27.0 (20 October) and 0.28.0 (27 October). The model page still dates ink-2 and ink-preview 17 September 2026 (30). In the 90 days from 3 August the SDKs released 4.0.0, 4.0.1, 4.1.0 and 4.2.0 (14 August to 2 September) and the MCP server several releases, with changelog entries for August and September (10). The deprecations page gives dated sunsets, and Sonic-2, Sonic Turbo and sonic-3-2025-10-27 were sunset after 20 October 2026, announced in the July changelog. No minimum notice period is stated and no Ink model has been retired, read as cartesia-tts is on the same page (5 of 10). The changelog has speech-to-text entries in May, June, July and September 2026, dated by month only, and no October entry (11 of 15). The Python SDK repository had 4 open issues on 9 October, per that check, not re-read on 1 November since GitHub's pages answered 403 (5 of 10). Official cartesia and @cartesia/cartesia-js SDKs at 4.2.0 (2 September 2026), generated for API version 2026-08-14 (15). CI workflows, release automation and lock files in both SDK repositories, with dependency updates merged on 23 October (8 of 10). 30 + 10 + 5 + 11 + 5 + 15 + 8 = 84.

Giggy Speech to Text 69

Read as a model, on the lines deepgram-stt and cartesia-ink-stt use (recency 30, dated releases 10, deprecation notice 10, support 15, issues 10, SDKs 15, health 10). The Python and Node SDKs released 0.2.1 and the Pipecat adapter 0.3.3 on 1 November 2026, and the examples and MCP repositories changed on 3 November (30). Since 7 October the Node SDK has four tagged releases, the Python SDK six and the Pipecat adapter six, the STT ones from 12 October (10). No deprecation policy or dated notice. The legacy https://giggy.ai/v1 origin and older SDK transports are kept 'for existing integrations' with no end date (2 of 10). No service changelog; the support and contact pages drew by script, and GitHub issues is the support URL in the plugin manifest (5 of 15). GitHub issues and reply times could not be read, since the GitHub API was not available to this session (0 of 10). Official SDKs current, giggy and @giggy-ai/sdk at 0.2.1 on PyPI and npm (15). CI workflows in five repositories and publish workflows, the n8n node through npm OIDC with provenance, and two runtime dependencies in each SDK; run results not seen (7 of 10). 30+10+2+5+0+15+7 = 69.

Cartesia Ink 47

Model reading. Standard keys (sk_car_...) and admin keys are separate types that are not interchangeable, and a server can mint an access token with an stt grant that lasts at most one hour. That token travels in the access_token query parameter on WebSockets, the API key itself goes in a header, and no key limited to speech-to-text was found. The STT WebSocket definition, re-read on 1 November, lists access_token in the query as a security scheme. The checklist takes 10 off for a secret that can travel in a URL query string as a documented option, and this run takes the full 10 for short-lived OAuth tokens in ?access_token= on every Google API listing, so the earlier 5 is brought into line (26, less 10, 16 of 30). The terms, last revised 14 June 2024, say inputs, outputs and interactions may be used to train models unless otherwise agreed, with an opt-out form in the Playground's data controls (5 of 20). Zero data retention covers STT audio and transcripts and is limited to the Enterprise plan. No retention period for other plans was found, and the DPA gives 180 days for deletion after termination (5 of 15). Credit usage and key metadata are readable through admin endpoints. No audit log was found (7 of 15). security.txt is valid until 3 October 2027 with a contact and a policy link to the trust centre. The docs and pricing page state SOC 2 Type II, HIPAA and PCI-DSS. The trust centre is drawn by script and was not read, and no bug bounty was found (14 of 20). 16 + 5 + 5 + 7 + 14 = 47.

Giggy Speech to Text 57

Model reading of the checklist, with training and retention in place of the least-privilege and injection lines, as for deepgram-stt and cartesia-ink-stt. Keys prefixed giggy_sk_ are shown once on creation or rotation, rejected immediately after revocation, limited to ten per space and tied to the space that made them, with REST budgets that Workspace members can set. No scopes or expiry were found. The key travels in a header only, with no query-string option documented (22 of 30). The workspace privacy guide states 'Giggy does not use user data to train models'. The privacy policy and terms draw by script and could not be read to confirm it (18 of 20). Requests can opt into 12-hour content retention and EEA-adequate compute, and a guide gives Batch audio a seven-day window, but no retention period for transcripts was found and results can vanish after an unstated 'retention cleanup' (9 of 15). Generation history per space and a request ID on errors; no audit log of key or admin actions was found (8 of 15). /.well-known/security.txt answered a client error, none of the six repositories has a SECURITY.md, and no disclosure policy, bug bounty or certification was found on the pages that rendered; the support page that may name a reporting route drew by script (0 of 20). 22+18+9+8+0 = 57.

Cartesia Ink 74

API reading, as for the other speech-to-text listings. Turn events carry a cumulative transcript that is never revised, so an agent reads one turn.end a turn, and word timestamps are opt-in on the manual and batch endpoints (20 of 25). There is nothing to page through. Batch takes a file of any length without client chunking, keyterms steer vocabulary, and there is no diarisation or subtitle output (12 of 20). Errors are structured JSON with error_code, title, message, request_id and an optional doc_url on API version 2026-03-01 and later (17 of 20). Transcription holds no state to duplicate and the SDKs retry failed HTTP calls. A realtime session cannot be resumed, and nothing on reconnecting was found (12 of 20). Realtime needs model, encoding, sample_rate and the version, turn settings have documented defaults, and there are official Python and TypeScript SDKs. Realtime accepts raw PCM only (13 of 15).

Giggy Speech to Text 68

API reading, as for the other speech-to-text listings. The transcript is compact, text, detected languages, audio duration and model metadata, with no word array to trim; on an MCP reading five tools would score 25 (20 of 25). Nothing to page through and no output options, no timestamps, subtitles or speaker labels, and recorded input must be PCM16 mono 16 kHz WAV, so the caller converts everything else; a 12-hour recording goes up in one request (10 of 20). Errors are structured with a type, code, message, status, request ID and parameter, and the reference says what each status means and what to do after a 504, but no list of code values was found and the SDK error classes expose only the status and the raw body (13 of 20). Upload retries are left to the caller with clear rules, read the generation first and never re-upload after a timeout; REST has no idempotency key and client_request_id is documented as no guarantee, while MCP transcribe_audio takes an idempotency_key. The SDKs never retry, and closing an unfinished live socket cancels it (13 of 20). Only the audio is required, with official Python and Node SDKs, but raw REST and MCP default to paid Fast while the SDKs default to free Batch (12 of 15). 20+10+13+13+12 = 68.