Smallest.ai Instant Voice Clone vs VoiceStudio

Smallest.ai Instant Voice Clone 72

Closed service with clear terms naming Smallest, Inc. under California law, effective 21 October 2026 (15 of 30). The Terms rule out training on customer content, the Privacy Policy (18 September 2026) gives retention criteria but no periods, and the Zero Data Retention page is specific about Enterprise options, with nothing on how long cloning samples are kept (16 of 30). A deprecation page with dated statuses (deprecated, rejects, removed, scheduled, retiring) and migration targets (16 of 20). Two regions named with locations. No sub-processor list was found and a DPA page exists unread (10 of 20). 15+16+16+10 = 57.

VoiceStudio 49

AGPL-3.0 source for the desktop app, with the Pro build under a separate commercial licence (30). Desktop retention settings and the website privacy notice (last updated 4 November 2026) are documented. The notice says the Pro desktop app sends usage and error metadata to Supermetry with no switch to turn it off, while the open-source app's analytics stay opt-in (15). A dated notice that Tauri 0.5.3 is the final Tauri release, with a migration guide, and no stated notice period (5). Open-source desktop analytics are opt-in and disclosed in docs/electron-privacy.md (15). 30+15+5+15 = 65.

Smallest.ai Instant Voice Clone 72

Read as a model, on the lines the other hosted cloning listings use (recency 30, dated releases 10, deprecation notice 10, support 15, SDKs 15, health 10). The models changelog has entries dated 29 October 2026 (30). Entries on 21, 24, 26, 28 and 29 October and in September (10). The deprecation page dates each retirement, such as lightning-v3.1 and its clones on 30 November 2026 with steps to migrate, though when the clone retirement was first announced wasn't found (7 of 10). Public changelog, support@smallest.ai and a Discord server, with no evidence read of how they answer (10 of 15). smallestai on PyPI released 5.12.3 to 5.14.0 between 1 and 7 October 2026, from a repository under fern-demo. The npm package is stale (10 of 15). Package repositories weren't cloned (5 of 10). 30+10+7+10+10+5 = 72.

VoiceStudio 50

Version 0.5.7 released on 26 October 2026, 10 days before this check (30). Four stable releases since 7 August, 0.5.4 (20 September), 0.5.5 (22 September), 0.5.6 (23 September) and 0.5.7 (26 October). The three 0.5.7 betas don't count (20). Issue and pull request replies couldn't be read, since GitHub's pages are closed to us (0, unchecked). No official MCP registry entry was found on a readable page (0). Package health not established. CI is public and 0.5.7 updated bundled dependencies past published advisories, but the CI results couldn't be read (0). 30+20+0+0+0 = 50.

Smallest.ai Instant Voice Clone 65

Graded on the hosted cloning API with the hosted lines. status.smallest.ai is the vendor's own status page with components per region and a dated incident feed (20). The feed, read on 6 November, has dozens of automated outage alerts on the API components in the 90 days from 8 August, most a few minutes long, and a cluster across every API region from 13 to 17 September. The 60-day uptime bars read 99.790 per cent for API USA and 99.849 per cent for API Global. No single outage of an hour was found, but repeated wide outages read as one major episode (10). Rate limits are published for TTS (one concurrent request, 100 a minute on pay as you go), while the per-organisation clone limit has no number (10 of 15). The limits page says to handle 429 gracefully and the clone errors say to retry a timeout, with no Retry-After, backoff figures or idempotency key (5 of 15). The pricing table lists an Enterprise SLA of 99.99 per cent (10). Instant cloning carries no beta label (10). 20+10+10+5+10+10 = 65.

VoiceStudio 45

Read with the local-package rubric, as a desktop app with a local MCP server over Streamable HTTP or a stdio shim. Installers for Windows, macOS, Linux, Android and iPhone on each GitHub release, with platforms and runtime setup documented (20). A public GitHub Actions workflow runs pytest on the backend and tests, the frontend tests and Electron tests on every pull request and push to main. Its results couldn't be read, since GitHub's pages are closed to us (15). Open crash and regression issues couldn't be read for the same reason (0, unchecked). Dated 0.x releases in CHANGELOG.md with Changed, Added, Removed and Fixed sections and upgrade notes, but no semver promise at 0.5 (10, the earlier partial, kept). Version 0.5.7, not 1.0 and not declared stable (0). 20+15+0+10+0 = 45.

Smallest.ai Instant Voice Clone 70

The API reference pages are generated with typed fields, enums and error schemas, but no OpenAPI file was found linked from the docs or llms.txt (5 of 25). llms.txt and a Markdown copy of every page (10). The guide says when to use the Pro pool, how to check modelIds before synthesis and why a mismatched clone returns empty audio, while the reference still names the retiring default (14 of 20). model and language are enums validated by the server, with required displayName and file, a 5 MB limit and MIME types listed, and tags as a comma-separated string (12 of 15). Examples in curl, Python and Node.js and a table of seven error cases with causes (14 of 15). Paths under /v1, a deprecation page with dated statuses and a models changelog with entries on 21, 24, 26, 28 and 29 October 2026 (15). 5+10+14+12+14+15 = 70.

VoiceStudio 60

docs/mcp.md is a Markdown guide in the repository, and voicestudio.sh says its docs also ship as Markdown and llms.txt (10). The guide's tool table names nine tools with a line on what each does, and the tool docstrings say what each returns and when a path or base64 input applies (15). Parameters are typed in the tool signatures and documented in the guide (format one of wav, ogg or opus, output modes, base path rules), but no JSON Schema is published on a docs page, so the contract line scores 0 (10 for typed inputs). Client configuration examples for Claude Code, Cursor and Codex CLI and documented refusals and error objects (10). CHANGELOG.md with dated releases, a Tauri to Electron migration guide and a supported-versions table (15). 0+10+15+10+10+15 = 60.

Smallest.ai Instant Voice Clone 20

No x402, MPP or L402 (0). Per-character TTS prices are public, about $0.175 per 10,000 characters on Lightning v3.1, but no price per clone is published and the models pricing table marks voice cloning No on pay as you go, against the docs (10). The agents pricing page states $10 in free credits, and whether a card is needed wasn't established (10 of 20). A person signs up in a browser and creates the key in the dashboard (0). 0+10+10+0 = 20.

VoiceStudio 30

No machine payment protocol (0). The Pro page lists Pro at $99 per user a year at launch ($199 list) and Lifetime at $199 one-time ($499 list), plan prices only, with purchases marked as coming soon (10). The AGPL desktop app is free to download and run with no account (20). Pro needs an account and a checkout, and the 7-day Pro trial needs an account (0). 0+10+20+0 = 30.

Smallest.ai Instant Voice Clone 62

The clone response is a compact object with the voice ID, status and signed URLs to sample clips rather than inline audio (20 of 25). The list call returns every clone on the organisation with no pagination or filter found (10 of 20). Errors carry a message and, on 500, an error_code enum with a stated fix, such as retrying voice_clone_timeout (16 of 20). No idempotency key, and retrying a timeout can leave a duplicate clone, with deletion through an oddly named route (4 of 20). Two required fields. The Python SDK smallestai 5.14.0 is current, while the npm package smallestai was last published on 16 April 2025 (12 of 15). 20+10+16+4+12 = 62.

VoiceStudio 70

Nine tools (generate_speech, clone_voice, describe_voice, design_voice, transcribe, list_voices, list_personalities, list_languages, check_health) with docstrings of about 4,600 characters in all, read from the server source on 5 November 2026 (25, ten or fewer compact tools). Output modes keep audio out of context (OMNIVOICE_MCP_OUTPUT_MODE=files returns a URL and path), and recent history returns at most 20 items (20). A missing voice returns a JSON object with an error message, and path arguments without a base path are refused with a reason (10, the earlier partial, kept). File inputs are confined to the base path, but no tool carries readOnlyHint or destructiveHint annotations and no retry guarantee is documented (5). Client configuration examples and sensible defaults (WAV output, agent-bound voice), no official SDKs (10). 25+20+10+5+10 = 70.