MiniMax Speech (T2A) vs Soniox Text-to-Speech

MiniMax Speech (T2A) 36

Model reading of the checklist, with training and retention in place of the least-privilege and injection lines. Plain API keys with no scopes found, and leaked keys may be disabled (20 of 30). The terms (effective 30 March 2026), as read for minimax-voice-cloning on 4 November, let MiniMax use input and generated content to operate, maintain, develop and improve its services, with no opt-out found, and the privacy policy rules out only training to profile or target consumers (5 of 20). The speech docs call every interface stateless, and no retention period for submitted text or generated audio was found (5 of 15). Responses carry usage_characters and a trace ID, with no audit log (6 of 15). No security.txt, disclosure policy, bug bounty or certification found (0 of 20). Total 36.

Soniox Text-to-Speech 87

Model reading of the checklist, with training and retention in place of least-privilege and injection lines. Project API keys now carry permissions per resource (TTS real-time, cloned voices, usage, model listing, temporary keys, with read and write levels), a missing permission returns 403 permission_denied, and an organisation API key can list, create and delete project keys. Temporary keys for browsers expire and are scoped to a usage_type (30; was 25). Soniox says audio and text are never used to improve its models (20). Nothing is stored unless you ask, and logs carry no content (15). Per-request usage logs record model, duration, tokens and the client reference id, readable through GET /v1/usage-logs, as soniox-stt reads them (10 of 15; was 5). SOC 2 Type 2, ISO/IEC 27001:2022 and HIPAA listed on the security page. The Python SDK's SECURITY.md (since May 2026) gives a disclosure route by email, GitHub, a maintainer or Discord. soniox.com/.well-known/security.txt answers 404 and no bug bounty was found (12 of 20; was 10). 30 + 20 + 15 + 10 + 12 = 87.

MiniMax Speech (T2A) 38

Read as a model, as minimax-voice-cloning is. The newest dated change touching speech is the MCP servers' August update, minimax-mcp-js 0.0.18 on 21 August 2026, 77 days before 6 November, and Speech 2.8 dates from 23 January 2026 (20 of 30). No dated API entries in the last 90 days (0). Support by email and Discord, not tested (6 of 15). No official SDK for speech; the MCP servers expose text_to_audio and list_voices (8 of 15). The Python MCP server has tests and moved to version 2 of the MCP SDK in August, with no CI workflow (4 of 10). Total 38.

Soniox Text-to-Speech 63

Read as a model, on the lines deepgram-tts 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 newest model change is tts-rt-v2 GA on 11 August 2026 and the newest dated vendor release Python SDK 2.10.0 on 2 October, 31 days before 2 November (20). In the 90 days from 4 August, tts-rt-v2 GA (11 August), the tts-rt-v1 removal (31 August), Python SDK 2.9.0 (11 August) and 2.10.0 (2 October), and the JavaScript SDK tag v2026.08.11 (10 August, counted half) (10; was 0). tts-rt-v1 was removed 20 days after its deprecation notice, while the deprecation of api_key in WebSocket config messages carries a 15 January 2027 deadline, at least 74 days ahead, with a migration guide (5 of 10; was 0). The models page and a support channel, untested (5 of 15). SDK issues couldn't be read because GitHub's pages answered 403 (3 of 10). SDKs for Python, Node, the web and React (15). The Python SDK has a test suite and lock file, but its CI file sits in .github/ rather than .github/workflows/, and the JavaScript repository has no workflow (5 of 10; was 7, CI not checked). 20 + 10 + 5 + 5 + 3 + 15 + 5 = 63.

MiniMax Speech (T2A) 56

API reading of the checklist. Audio comes back hex-encoded inside JSON by default, with subtitles opt-in and format, bitrate and sample rate settable (15 of 25). Voices are listed by type through POST /v1/get_voice, which has no pagination and returns 300+ system voices for all (10 of 20). Errors come as numeric codes in base_resp with a short remedy for each, inside responses documented only as 200 (13 of 20). No idempotency key, and the MCP tools carry no read-only or destructive annotations (10 of 20). Two required fields. No official SDK for speech, with MCP servers in Python and JavaScript (8 of 15). Total 56.

Soniox Text-to-Speech 78

API reading of the checklist. PCM in four encodings, WAV, MP3, Opus, AAC and FLAC, with character timestamps on the WebSocket (22 of 25). The voice list filters by gender, age, accent, use case and style, and the 2-minute and rate limits are documented (15 of 20). Machine-readable error_type values an agent can branch on (20). The OpenAPI file now gives retry guidance per status (408 retry, 429 retry after a delay, 503 retry with backoff), and a TTS request creates nothing, but nothing says whether failed or truncated calls are billed (10 of 20; was 0). Model, language, voice, format and text are all set per request. SDKs for Python, Node, the web and React (11 of 15). 22 + 15 + 20 + 10 + 11 = 78.

MiniMax Speech (T2A) 75

Atlassian Statuspage at status.minimax.io with a Text-to-Speech component (20). Its history, re-read on 6 November, shows three incidents in the 90 days from 8 August, all titled elevated errors and labelled major, 22 minutes on 28 September, 3 minutes on 16 October and 2 minutes on 19 October, and none in November. Elevated errors with no fuller scope are read as partial, as on minimax-api and minimax-voice-cloning, so minor only (20). T2A is limited to 60 requests a minute, with higher limits on audio subscriptions (15). Error codes 1002 and 2045 say to retry later and avoid sudden increases, with no Retry-After or backoff figures (10 of 15). No SLA (0). No beta or preview label on T2A (10). 20 + 20 + 15 + 10 + 0 + 10 = 75.

Soniox Text-to-Speech 87

Instatus page at status.soniox.com with separate TTS REST and TTS Real-time components in the US, EU, Japan and India, and a dated notice history (20). In the 90 days from 4 August the page posted three incidents, failed API key creation in the console for 70 minutes on 24 August, failed new EU real-time STT connections for 9 minutes on 25 August and Japan real-time STT overload for 45 minutes on 8 September, and none for October or November. None touches TTS, so the record is clean for this listing (30). Limits published, 100 requests a minute and 3 concurrent requests or streams, raisable in the console (15). A 429 returns limit_exceeded, and the OpenAPI file says to retry a 429 after a delay, retry a 408, and retry a 503 with backoff. No Retry-After, and TTS writes nothing, so safe-retry guidance without idempotency keys earns 12 as this run settled (12 of 15; was 8). No SLA found (0). tts-rt-v2 is GA in all four regions (10). 20 + 30 + 15 + 12 + 0 + 10 = 87.

MiniMax Speech (T2A) 65

Closed API with clear terms under Singapore law naming Nanonoble Pte. Ltd., and MIT MCP servers (15 of 30). Terms and privacy policy agree that inputs may be used to improve services and that data sits in a United States data centre, with no retention period, as on minimax-voice-cloning (8 of 30). The API release notes carry a dated legacy notice of 20 June 2025 for the old text-to-voice API with service continuing, and older speech models stay accepted, with no written policy or notice period (15 of 20). Data location stated as the United States and an EU representative named, third parties only by category (10 of 20). Total 48.

Soniox Text-to-Speech 73

Closed service under terms updated 2026-06-29, Python SDK under Apache-2.0 (15). The security page, the terms and the TTS docs agree on no storage and no training, with async data deleted after 30 days (25 of 30). Model removals are dated on the models page with 20 days' notice for v1, and the WebSocket authentication change is dated to 15 January 2027 with a migration guide, but there is no written policy (12 of 20; was 10). Regional hosts in the US, EU, Japan and India and a data residency page, no sub-processor list found (12 of 20). 15 + 25 + 12 + 12 = 64.