v0.0.1-beta.1

Voicy speaks the sixteen languages the bot family shares, transcribes audio
files and video rather than only voice messages, and no longer holds a
recording in memory to do it.

Added

• Sixteen interface languages, matching the set searchy and vido offer. Strings
  moved out of Go into internal/i18n/translations.json, and the picker shows
  each language by its own name. Ukrainian and Belarusian used to collapse into
  Russian, which also meant Voicy overwrote a person's real choice in the shared
  Core language hub with a coarser one.
• Command descriptions are published per language through setMyCommands, so
  the command menu is no longer English for everyone.
• Audio files, video and documents are transcribed, not just voice messages and
  circles. In a direct chat, voice, circles and audio files are still handled
  without being asked; video and documents wait for an explicit /v, so a
  Deepgram call is never spent on a file shared for some other reason. A
  document is accepted on its MIME type or file name and refused with a plain
  answer when it carries no sound.
• ffmpeg reduces video, and any media above EXTRACT_ABOVE_BYTES, to a 16 kHz
  mono Opus track before Deepgram sees it. This is what Deepgram recommends for
  large video, and it keeps a large upload from crossing the network twice.
• FFMPEG_PATH (default ffmpeg, empty disables extraction),
  EXTRACT_ABOVE_BYTES (default 20 MB) and MEDIA_TMP_DIR.
• Statistics count files as their own category next to voice messages and
  circles.

Changed

• Media is streamed to a temporary file and read from there, instead of being
  buffered whole in memory. That buffering, not Deepgram, was what capped
  MAX_MEDIA_BYTES: with several transcriptions running at once the ceiling
  was memory. Deepgram's own limit is 2 GB, and the Deepgram request body is now
  streamed from disk too, so a retry can replay a body of any size.
• The Telegram client exposes DownloadToFile rather than returning bytes.

Migrations

• migrations/005_media_kinds.sql widens the transcript kind constraint to the
  new media types and adds user_stats.file_count.

Operations

• The image now installs ffmpeg. Without it, video is sent to Deepgram
  unextracted and may not transcribe at all; the bot logs a warning at startup.
• MAX_MEDIA_BYTES can be raised well beyond the previous memory-bound value
  when a self-hosted Bot API server is in use.

Container image

ghcr.io/freshlabdev/voicy@sha256:b17f4b5f2c4b9fc2979c1a6b92689e2f438e4dcec9460916e8650b4dc0e96ba0