Voice clone

log keeps one voice: matt's, cloned from the voice notes he sent himself. When he says "send <person> a voice note saying …", it goes out as the platform's own voice note in his voice (rule 142). A voice note only ever goes out when matt asks for one, never on the agent's own initiative, and never any other audio.

The voice layer is provider-agnostic (rule 143), the way proxies are: one interface in src/extensions/voiceClone, and each provider an extension of its own.

The provider is --provider, else LOG_VOICE_PROVIDER in .env.local, else inworld. A new provider is one more extension implementing VoiceProviderClient and one row in PROVIDER_CLIENTS.

What log can do with it

log fish clone|speak became log clone voice|speak and is refused with that form. On X, log twitter voice sends it as an X Chat voice message through vendor/twikit, the way X Chat's own app does (rule 151).

The clone

log clone voice finds matt's own recordings in what the engines kept: Discord voice messages (~/.logCli/discord/<account>/media/<chat>/<messageId>_voice-message.ogg) and Instagram voice clips (voice_<itemId>.m4a). It keeps only the ones he sent: fromMe in the conversation store, or his ownMessage in the .db. Nobody else's voice is ever sent to a provider. It takes them shortest first, each at least 5 s, within the provider's limit, converting them to wav with ffmpeg when the provider wants wav. Without --apply it only lists them. Each provider's clone is made once and kept in ~/.logCli/voice/clone.json (LOG_VOICE_CLONE_DIR moves it), keyed by provider, with the recordings it came from.

A voice note

Every voice command takes --text only: a path or an audio file name as the text is refused (voiceNoteInvalid), and there is no flag for audio. Without --live nothing is generated or sent. With --live --approvedBy:

1. the account catches up (rule 135) and the text goes through the voice gate every send runs (rule 137); a refusal there sends nothing;
2. the provider says it in the clone (as wav), and ffmpeg converts it: Discord ogg/opus 48 kHz mono, Instagram and Slack AAC in .m4a;
3. its duration and waveform are measured from the audio itself;
4. the chat is claimed (rule 133), the engine's voice write sends it through the account's one session, and the claim is released.

The engines take only a file log made under ~/.logCli/voiceNotes with the platform's suffix (voice_note_path). Instagram's upload goes on instagrapi's curl adapter under the one-call lock (rules 128, 129); Slack's on gabriel's curl_cffi transport (rule 125). Any refusal stops, and nothing is retried (rule 94).

Spec: docs/specs/voiceClone.md. Use case: Sending a voice note.

Sounding like his own voice notes, and saying what was written (rule 147)

Every spoken file goes through two steps in speakToFile before anything uses it:

1. Shaped by processLikeMatt (helpers/realism.ts). The settings in REALISM were measured from five of matt's own Discord voice notes against the clone: phone squash and his top end, his slow mic drift (about 1.9 dB), the tone swaying as the phone moves, phone-mic saturation on loud words, and his own room tone, cut from the quiet gaps in his own voice notes (quietPieces; a recording with a floor under -60 dB, like a generated one, gives none) and laid under the voice where his sits (30 to 33 dB under it). The same text gives the same file (seeded).
2. Listened back with AssemblyAI's word timings (listenTo, kept by the file's hash) and checked by readSpoken (helpers/spoken.ts). Words in SPOKEN_LEXICON are written the way the clone says them right before it speaks (speakableText: op goes as OP) and must come back as themselves, as one word under 400 ms. A lexicon word heard as something else or spelled out, or more than a quarter of the words different, is refused spokenMismatch. An acronym not in the lexicon is refused pronunciationUnknown before anything is spoken, and a note that can't be listened back is never sent (spokenCheckFailed).

A lexicon entry is added only when matt says how he says the word.

In another language (rule 148)

When matt asks for a voice note in another language, the translation goes out with --translationOf "<the original>" and --language <code> (zh-CN, es-ES ...). The voice gate grades the original, because it compares a draft with matt's own writing and can't match a translation to it. The clone speaks in that language instead of its own, and the listen-back check reads Chinese, Japanese and Korean one character at a time.