Repository navigation
v0.1.1-alpha.1
Pre-release
Pre-release
·
11 commits
to main
since this release
Voice messages in Arabic, Persian, Belarusian and 23 more languages stop coming
back as a wrong guess in Latin letters: Voicy names the language for the people
who speak it, and listens again with detection when that guess was wrong.
This alpha runs on the production bot. Voicy has no test bot (see
docs/versioning.md), so real Arabic and Persian traffic is the only way to
check the confidence bar below. It either earns a stable tag or is rolled back.
Checked by unit tests and by the database tests against PostgreSQL 17,
migration 007 included.
Added
- Voice messages in Arabic, Persian, Belarusian and 23 more languages.
Deepgram's automatic language detection knows 35 languages, and Arabic,
Persian and Belarusian are not among them: an Arabic voice message came back
labelled French and written in Latin letters, Hebrew as Polish, Kazakh as
Turkish. nova-3 transcribes all of them, but only when the request names the
language. Voicy now names it for people whose language is one of those, taking
it from their pick in Settings, then from the language the shared core
resolved for them, then from the language of their Telegram client. Everyone
else is detected as before. The full list isdeepgram.HintLanguages. - A file in another language than expected is not lost. If the named run
comes back empty or under 85% confidence (an Arabic speaker forwarding an
English voice message), Voicy runs the same file again with detection and
keeps the better transcript. That second run is billed, so it counts toward
USER_DAILY_AUDIO_LIMITandvoicy_audio_seconds_total; the new
voicy_language_fallbacks_totalcounts how often it happens. - Settings → Recognition language. Auto, or one of the languages above, for
people whose Telegram client and interface language differ from the language
they speak. Migration007addsuser_settings.stt_language, empty by
default. Texts in all sixteen interface languages.
Changed
- The transcript cache variant now ends in
-l<code>when a request names a
language (sf1-p1-fw0-pf0-d0-lar), so the same file read as Arabic and read as
detected French are two rows.settings.DefaultVariantand every row cached
before are unchanged. Files already cached for an Arabic or Persian speaker
are transcribed once more under the new variant, which is the point. voicy_audio_seconds_totalis counted where the Deepgram request is made
rather than before it, so a repeated run is in it.- Migration
007_stt_language.sqladdsuser_settings.stt_languageand is
applied on start; nothing to run by hand. v0.1.0 never reads the column, so
rolling back to it needs only the previous image.
Known Limitations
- The 85% bar was set on synthetic speech (macOS text-to-speech, plain, with
pink noise and in an 8 kHz telephone band), because no recorded voice messages
were available. Language detected correctly scored 0.86 to 1.0; audio in the
wrong language came back empty; the detector's wrong guesses scored 0.0 to
0.77. Re-check it againstvoicy_language_fallbacks_totalonce real Arabic
and Persian traffic arrives. - Uzbek, Khmer and Burmese are not nova-3 languages and are not covered.
- In a group, the language named is the requester's, not the author's of the
voice message; the repeat with detection covers a mismatch at the price of a
second run.
Container image
ghcr.io/freshlabdev/voicy@sha256:67095941e1ba62915a6ac7a89d39e6a14e0e4ea62de8a3fb3b40bfbde6bd8a16