Inworld

Inworld (API docs; subscribe through that link for $10 in credits) is the default voice provider for the voice clone (rule 143). Instant voice cloning is free on every plan, and the free plan includes about 70 minutes of speech.

The two calls

The client (src/extensions/inworld/client.ts) knows only these, each with Authorization: Basic <INWORLD_API_KEY> (the Base64 credentials from the Inworld portal, sent as they are):

• POST https://api.inworld.ai/voices/v1/voices:clone, JSON displayName, languageCode (LOG_VOICE_LANGUAGE, default en-IE; when Inworld answers 400 INVALID_ARGUMENT it is sent again as en-GB, Inworld's own example of a British accent), description naming the accent (LOG_VOICE_ACCENT, default Irish), and voiceSamples, each audioData as base64 wav with its transcription when ~/.logCli/transcripts holds matt's AssemblyAI transcript of that file (Inworld says TTS-2 cloning makes its own and ignores it, so a missing one is not paid for). log clone voice --apply --replace makes it again, keeping the old voice id under previous (rule 145). Inworld works best with 10 to 30 s and cuts longer clips at 30 s, so log sends matt's own voice notes shortest first up to 30 s in all. The answer's voice.voiceId ({workspace}__{voice}) is kept in ~/.logCli/voice/clone.json under inworld.
• POST https://api.inworld.ai/tts/v1/voice, JSON text (up to 2,000 characters; log keeps a voice note under 600), voiceId, modelId, language (the language the clone was made in) and audioConfig (audioEncoding WAV, speakingRate from LOG_VOICE_SPEED, default 1.15). The answer's audioContent is base64 audio. No accent instruction is sent: the clone carries the accent (rule 147). On inworld-tts-2 the text gets [breathe] at the start and once at least 8 words have gone by (withBreaths); emotion tags such as [excited] pass through as written. inworld-tts-2-flash ignores tags, so it gets none.

A clone made on Inworld speaks with inworld-tts-2 (its preferred model, rule 146) unless --model says otherwise.

Inworld's own docs say instant cloning may not do well on unique accents and gets closer the nearer the sample is to 30 s, so matt's clone is one continuous 30 s of his own Discord voice note rather than several short ones (rule 147).

Setting it up

• INWORLD_API_KEY in .env.local; its name is in .env.example.
• log clone voice, read the list, then log clone voice --apply once, and log clone speak --text "…" to hear it.