2.1.0

Words you never said

• Local Whisper models no longer add text out of silence. A pause made the model fall back on something memorized in training - "Thank you." in English, a subtitle-credit line in Russian, equivalents in Turkish, Czech and Arabic - and it landed in your transcript.
• Caught three ways: while decoding, again on the finished segments by reading confidence against silence, and by name for the phrases distinctive enough to match. It runs whether or not you have filler-word removal turned on - a filler is something you said and chose to keep, a hallucination is not.

Three realtime transcription providers

• Cartesia Ink 2, Gemini 3.5 Transcribe Live and OpenAI GPT Live Transcribe each transcribe over a socket while you speak.
• Each one falls back to a matching batch model when you import an audio file instead, so nothing changes about file transcription.

Models and providers

• New models for enhancement: GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash and GLM 5.3 Flash. Gemini 3.8 Flash becomes the default for both enhancement and Gemini transcription.
• Newer OpenAI reasoning models now work everywhere. The checks recognized only the GPT-5 generation by name, so anything after it was sent a setting it rejects and failed every call - including when reached through a gateway.
• Cerebras and Groq drop models their providers retired, and saved selections migrate automatically rather than snapping back to a default. Cerebras moves to Qwen 3.8 27B; Groq's Qwen 3.6 27B had an output cap below what a single enhancement needs, so every call failed with a rate-limit error no matter how slowly you worked.
• OpenCode Zen and Go are faster and hit rate limits far less often. Every request identified itself as an anonymous client, which cost both the request budget and the prompt cache between calls.

A Top Models card worth reading

• The card now recommends one model per provider instead of the ten highest-scoring rows. Half the old list was second-best models from providers already on it, and the same model showed up three times through three different routes.
• Nine providers now, four of them usable without paying at all.
• Every model in the ranking was re-measured from scratch and the published table grows to 59 models. Six OpenCode Go models appear for the first time - they had never been measurable. The full ranking has all of them.

Failures that used to be silent

• The OpenRouter model list survives a failed refresh. A single network error used to empty the picker until some later refresh happened to succeed.
• If you deliberately chose your provider's default model, VivaDicta no longer moves you off it. Automatic selection treated that as "nothing picked" and could switch you to whichever of roughly five hundred model ids sorted first.
• An enhancement that comes back empty, or that stops partway through, is now reported as an error instead of quietly replacing your dictation with nothing or with a truncated version of it.
• Checking an API key no longer reports a good key as invalid. The check ran against your currently selected model, so a retired model or one your plan cannot reach failed the key itself.
• The system information dump reports the Viva Mode that actually ran, along with its model, provider and language. A support report could previously describe a setup that had nothing to do with the recording.