Voice notes

Every voice note on X and Discord, sent to matt or by him, is transcribed with AssemblyAI and written into the conversation store, so search, the voice gate and SubQ see what was said. Rules 26, 41, 88, 89, 94, 105, 106 and 126.

AssemblyAI

REQ-TRANSCRIBE-01: Upload, ask, poll until done

A file is uploaded with POST https://api.assemblyai.com/v2/upload, its transcript is asked for with POST /v2/transcript naming the upload url, and the transcript is read with GET /v2/transcript/<id> every few seconds until AssemblyAI says completed or error, or 15 minutes pass. Every call carries the key in the authorization header.

Checked by: code and SubQ

REQ-TRANSCRIBE-02: No key means no call

The key is ASSEMBLYAI_API_KEY or ASSEMBLY_AI_KEY (the environment, the repo's .env.local, ~/.buzz/.env.local or ~/filmcraft/.env.local), never printed. Without it nothing is sent to AssemblyAI and nothing is read from a platform: log transcribe refuses with assemblyAiKeyMissing, and a read that met a voice note says it isn't transcribed.

Checked by: code and SubQ

REQ-TRANSCRIBE-03: AssemblyAI's failure is passed on

An answer that isn't 200, or a transcript whose status is error, comes back as transcriptionFailed with AssemblyAI's own words, and the run says which voice note failed.

Checked by: code and SubQ

REQ-TRANSCRIBE-04: A file is transcribed once

Each transcript is kept by the sha256 of its file in ~/.logCli/transcripts; a file whose hash is kept is answered from there and never sent to AssemblyAI again.

Checked by: code and SubQ

The store

REQ-TRANSCRIBE-05: The marker becomes what was said

A message's (audio: <name>) marker is replaced by (voice note: "<transcript>"), and the leftover (audio saved: …) marker goes, only for a message whose platform, account, message id and file name are tied to a kept transcript.

Checked by: code and SubQ

REQ-TRANSCRIBE-06: A fresh read never loses a transcript

Saving conversations (an export, twitterDms, searchIndex) writes the kept transcripts back in before anything is written, so a conversation read again keeps its voice notes as words.

Checked by: code and SubQ

REQ-TRANSCRIBE-07: Discord keeps its attachments

Discord's export, read and watch keep each message's attachments (name, type, duration), and an audio attachment adds its (audio: <name>) marker to the message's text, so a voice message is no longer an empty message.

Checked by: code and SubQ

Finding and fetching

REQ-TRANSCRIBE-08: X voice notes come through the account's one session

An X voice note's file is the one the X Chat read saved; a missing one is fetched by reading its conversation again through the account's engine with its audio events reopened, never by a request of log's own.

Checked by: code and SubQ

REQ-TRANSCRIBE-09: Discord voice notes come from Discord's sound search

Discord voice notes are found with Discord's own message search for sound (has=sound), one chat at a time, paused between calls, and each file is fetched from its fresh attachment url through the vendored client.

Checked by: code and SubQ

REQ-TRANSCRIBE-10: A refusal stops it

Any refusal from Discord's search (a 429, a 401, any answer that isn't 200 or 202) stops the whole search with Discord's answer, keeping what was found; nothing is retried. An X read that fails stops the fetching for that account.

Checked by: code and SubQ

New voice notes

REQ-TRANSCRIBE-11: Reads and the inbox transcribe as they go

discordRead, slackRead, twitterDms with or conversation, and a voice message the Discord watch hands to the inbox are transcribed before they are shown; the inbox hook shows a kept transcript and says when one isn't transcribed yet.

Checked by: code and SubQ

Every platform

REQ-TRANSCRIBE-12: Audio is detected once, for every platform

One shared detector, in Python (python/platformKit/media.py) and TypeScript (src/platformKit/media), says whether an attachment is audio: a content type of audio/* or video/ogg, an extension (.ogg .oga .opus .m4a .mp3 .wav .aac .webm), or the platform's own flag (Discord's voice-message flag 8192 with a duration, or a waveform; X Chat's audio media; Slack's slack_audio file subtype). A video or image type is never audio. The Discord, Slack and X engines and the transcribe path all use it; none has a list of its own.

Checked by: code and SubQ

REQ-TRANSCRIBE-13: A new platform keeps its attachments from the start

log new platform writes a message shape into the new engine that keeps each message's attachments and marks audio with the shared detector, and check:platforms refuses any platform engine that answers with messages but doesn't import the shared detector or keep attachments.

Checked by: code and SubQ