IBM Watson Speech to Text vs OpenAI Speech to Text

IBM Watson Speech to Text 33

Read as a model. The newest release note is 29 June 2026 (parameter_set for large speech models), 130 days ago (10). No dated entries in the 90 days from 8 August (0). Release notes, IBM support and a public docs repository last changed 30 June 2026 (8 of 15). Official SDKs, ibm-watson 11.2.0 on PyPI (26 January 2026) and 12.2.0 on npm (January 2026) (10 of 15). CI not read (5 of 10). Total 33.

OpenAI Speech to Text 71

Model reading. The last dated change to transcription is still the deprecation notice of 26 August 2026, 67 days before 1 November, after the release of gpt-transcribe and gpt-live-transcribe on 28 July (20 of 30). The deprecations page promises at least six months' notice for generally available models, and the four transcription models got six months to the day. On 14 October GA GPT-4 series models got 92 days with no reason stated, so the policy reads as a minimum of two to three months in practice, as the OpenAI API and Decisions API listings read it (8 of 12, was 12). Four of the five file transcription models were deprecated within a month of their replacements shipping, and the replacements don't cover diarisation, timestamps or subtitles (2 of 8). The dated changelog has two transcription entries in 90 days among dozens for the API (13 of 15). Python SDK replies were not sampled (5 of 10). Python SDK v3.42.0 was tagged on 31 October 2026, with v3.41.0 on 28 October and v3.40.0 on 27 October (15). CI, CodeQL and a breaking-change check run in the repository (8 of 10). 20 + 8 + 2 + 13 + 5 + 15 + 8 = 71.

IBM Watson Speech to Text 55

No public OpenAPI or other machine-readable definition was found (0). No llms.txt (0). The reference says which parameters each model family supports and the docs cover each feature (15 of 20). Typed parameters with defaults, ranges and enums (13 of 15). curl and SDK examples on every method, and errors with error, code, code_description and warnings, with invalid arguments reported as warnings (12 of 15). Versioned /v1 paths and dated release notes back to 2015 (15). Total 55.

OpenAI Speech to Text 89

Model reading. The public OpenAPI 3.1 document in openai/openai-openapi (version 2.3.0, last committed 31 October) covers /audio/transcriptions and /audio/translations and names gpt-transcribe (25). llms.txt and a Markdown twin of every docs page (10). The transcription overview says which model to use for files, live audio, speaker labels, timestamps and translation, and the file guide now opens by recommending gpt-transcribe and saying the specialised examples use deprecated models (18 of 20). response_format, timestamp_granularities and model are enums in the document. language is a plain string, and which formats each model accepts is stated in prose only (12 of 15). Examples in seven languages and curl, and an error codes page with a cause and fix for each. The Markdown twin of the endpoint reference now lists the body parameters, including languages and keywords for gpt-transcribe, but its first example still names the deprecated gpt-4o-transcribe, and the guide lists seven input formats where the document lists nine (12 of 15, was 11). A dated changelog and deprecations page. gpt-transcribe has one alias and no dated snapshot to pin (12 of 15). 25 + 10 + 18 + 12 + 12 + 12 = 89.

IBM Watson Speech to Text 68

Model reading of the checklist, with training and retention in place of the least-privilege and injection lines. IAM API keys per instance with Reader, Writer and Manager roles and short-lived bearer tokens (30). The WebSocket interface takes the token as an access_token query parameter, so 10 off (20 of 30). Request and response data are logged by default to improve the service, and any account can opt out per account or per request (10 of 20). Async results expire after a week by default (results_ttl shortens it), a delete method exists, and DELETE /v1/user_data removes data by customer ID (10 of 15). Activity Tracker events record changes to the service's state (10 of 15). Valid security.txt to 6 December 2026, PSIRT and HackerOne, HIPAA readiness on Premium. Certifications weren't read (18 of 20). Total 68.

OpenAI Speech to Text 97

Model reading, on the lines the OpenAI API and Decisions API listings use. Project keys with permissions per key, service-account keys through the Administration API and workload identity federation, per the Decisions API listing's read of 8 October. The access-control guide, re-read on 1 November, puts audio under one Model capabilities permission with chat completions, embeddings and images, so a restricted key can't be limited to the audio endpoints alone (27 of 30, as the Decisions API listing, was 26). The data controls page says API data isn't used for training unless the customer opts in (20). Both audio endpoints are listed with no abuse-monitoring retention, no application state and Zero Data Retention eligibility, re-read on 1 November (15). The Admin API has an audit log endpoint, and the OpenAPI document has GET /organization/usage/audio_transcriptions grouped by project, API key or model, the per-key usage view both other OpenAI listings got 15 for (15, was 13). security.txt, re-read on 1 November, is PGP-signed, names a Bugcrowd programme and a coordinated disclosure policy and has no Expires field. trust.openai.com, read on 1 November, names SOC 2 Type 2 and ISO/IEC 27001, 27017, 27018 and 27701 (20, was 12 when no trust centre was read, the award both other OpenAI listings carry). 27 + 20 + 15 + 15 + 20 = 97.

IBM Watson Speech to Text 40

IBM Cloud has a status page at cloud.ibm.com/status, but it is drawn by script and showed no components or history to our reader, read as gemini-live's script-shell page was (10 of 20, and 5 for no readable history). Plus allows up to 100 concurrent transcription requests across WebSocket and HTTP (15). No 429 or backoff guidance was found in the API reference or the docs repository (0). No SLA was read for the service (0, unchecked). Next-generation and large speech models are GA, language by language (10). 10 + 5 + 15 + 0 + 0 + 10 = 40.

OpenAI Speech to Text 65

Hosted reading. status.openai.com (incident.io) lists Audio and Realtime as components of the API group (20). The feed, read on 1 November, names the Audio component in three incidents in the 90 days from 3 August, and their incident pages give the Audio impact. 17 September, "elevated error rates across API models", Audio degraded from 20:20 to 21:20 UTC with ten other API components, an hour. 29 September (last updated 6 October in the feed, which the 9 October check dated 6 October), errors across ChatGPT, Codex and the API, Audio degraded from 17:52 to 23:14 UTC, 5 hours 21 minutes. 21 October, degraded performance across 33 components, Audio from 02:26 to 02:48 UTC, 22 minutes and minor. The first two are API-wide and an hour or more, so two majors, and two majors give 5 of 30, as the OpenAI API, OpenAI Moderation and OpenAI Image API listings read the same two incidents (5, was 20 when the durations were unread). The July incidents are outside the window. The gpt-transcribe model page, re-read on 1 November, gives 5,000 requests a minute on Build, 10,000 on Launch and 30,000 on Grow (15). The rate limits guide, re-read on 1 November, documents Retry-After on 429 and 503, the slow_down and server_is_overloaded codes and backoff with jitter, and transcription is a stateless call (15). No SLA covering the audio endpoints was found, read as for the OpenAI Decisions API listing; the Scale Tier pages are on openai.com and weren't requested (0). gpt-transcribe carries no preview label (10). 20 + 5 + 15 + 15 + 0 + 10 = 65.

IBM Watson Speech to Text 66

API reading of the checklist. Timestamps, word confidence, alternatives and speaker labels are opt-in (15 of 25). Async jobs and custom models list, and async results can go to a callback URL (12 of 20). Structured errors and warnings an agent can act on (15 of 20). Async jobs take a user_token, results have a TTL, and jobs can be deleted, with no idempotency key (12 of 20). Most parameters are optional. Official SDKs including Python, Node and Swift, though the Python and Node SDKs last released in January 2026 (12 of 15). Total 66.

OpenAI Speech to Text 78

API reading, as for the other speech-to-text listings. The default json response is the text, detected languages and usage, text returns a bare string, and segments, words and speaker labels come only when asked for on the models that have them (20 of 25). There is no transcript store to page through. Files stop at 25 MB, the caller splits longer audio, and the gpt-transcribe model page marks the Batch API as not supported (10 of 20). Errors carry message, type, param and code, and the error codes page gives a fix for each status (18 of 20). The call is stateless and safe to retry, and the Python SDK retries 408, 409, 429 and 5xx twice by default. No idempotency key applies to this endpoint (15 of 20). file and model are the only required fields, with guide examples for JavaScript, Python, Go, Java, C# and Ruby SDKs and a CLI (15).

IBM Watson Speech to Text 40

No x402, MPP or L402 (0). Per-minute Plus prices published in the docs without a login (20). The Lite plan gives 500 free minutes a month, and no page read says a card is needed (20). A person creates an IBM Cloud account and service instance in a browser (0). Total 40.

OpenAI Speech to Text 30

No x402, MPP or L402 on the pricing page, re-read on 1 November (0). $0.0045 an audio minute for gpt-transcribe, $0.017 for gpt-live-transcribe and $0.006 for whisper-1 on the public pricing page (20). The rate limits guide lists a Free tier capped at $100 a month in allowed countries, while the gpt-transcribe model page's rate-limit table starts at Build, which needs $5 of credit. Half, the reading both other OpenAI listings use for the same pages (10, was 0). A person signs up in a browser and creates the key in the console (0). 0 + 20 + 10 + 0 = 30.