Repository navigation
Releases
v0.1.3
The GitHub organization FreshLabDev is now asterfield-dev, and the bot moves
with it. Nothing a person sees in Telegram changes.
Security
- Built with Go 1.26.11, which fixes GO-2026-6859 in
net/http.
govulncheckfinds the bot reaching it through its HTTP client and health
server.
Changed
- Images are published as
ghcr.io/asterfield-dev/voicy. The organization
was renamed, and GitHub Container Registry does not redirect the old name:
ghcr.io/freshlabdev/voicyanswers 403. Point the deployment's image at the
new path; every earlier digest is there under it too. - tg moves from
v0.0.1-alpha.7tov0.4.0-alpha.2, whose module path is
github.com/asterfield-dev/tg. voicy needed no change for it. - Links, the module path and the docs point at
github.com/asterfield-dev.
Old links still redirect.
Known Limitations
- Released as a stable patch without a pre-release first, at the owner's
request: the change is the organization rename and the tg upgrade, checked
by the full test suite, not by a run on a test bot. v0.1.2was tagged but never published: its vulnerability scan stopped
the release on GO-2026-6859. This version is that one plus the Go update.
Container image
ghcr.io/asterfield-dev/voicy@sha256:70bb60c394848ace18142955c327c6a209b8fc7e61875a5a1d13c00b0c4e3376
v0.1.1
Voice messages in Arabic, Persian, Belarusian and 23 more languages that
Deepgram cannot detect are transcribed in the speaker's language instead of a
wrong guess in Latin letters.
This is v0.1.1-alpha.1 promoted unchanged. It has run on the production bot
since 18 October without an error, but in those two days it handled a single
voice message, in Russian: the named-language path and its 85% confidence bar
have not yet met real Arabic or Persian traffic. It is promoted on the unit
tests and the database tests against PostgreSQL 17. If that path misbehaves,
the fix ships as a patch.
Container image
ghcr.io/freshlabdev/voicy@sha256:4b782773fa5ab5abfc608a1e654dcee3e3d1f7858e56707e2be88de6231c0cf7
v0.1.1-alpha.1
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
v0.1.0
Voicy goes public with a ceiling on what one person can spend: an hour of new
audio in any 24 hours, with cached transcripts still free. The build moves to
Go 1.26.9, and the German and Ukrainian texts now read correctly.
No pre-release of this version ran anywhere. Voicy has no test bot, because a
second bot needs its own Deepgram budget (see docs/versioning.md), so a
pre-release would have gone to the same production bot this release does. The
allowance was checked by unit tests and by the database tests against
PostgreSQL 17, migration 006 included.
Added
- A daily allowance per person. Deepgram bills every second it hears, and
nothing stopped one person from forwarding podcasts all day: audio files are
transcribed in a direct chat without being asked, up to an hour each. Now one
person gets at mostUSER_DAILY_AUDIO_LIMITof new audio in any 24 hours,
one hour by default. The window rolls instead of resetting at a midnight that
would fall at a different hour for everyone. A file that does not fit is
refused before it is downloaded, and the answer says how large the allowance
is and how long until this file would fit. A file already in the cache costs
nothing and is still served./vprefusals stay private. voicy_daily_limit_refusals_totalcounts the refusals.
Fixed
- German: the start screen and help said "ein Videokreis" where the accusative
"einen Videokreis" belongs, the first sentence every German reader sees. - Ukrainian: "Не розібрав мову" read as "did not understand the language"; the
empty-speech answer now says "мовлення", speech.
Security
- Built with Go 1.26.9. Under Go 1.26.6 Voicy's code called into nine
standard-library vulnerabilities innet/http,net/textprotoand
crypto/tls(GO-2026-6603, -6605, -6607, -6608, -6610, -6611, -6612, -6613,
-6617).golang.org/x/textmoves from v0.39.0 to v0.43.0 for GO-2026-6629,
which pullsgolang.org/x/syncto v0.24.0. Voicy publishes no port outside
its Docker network, so none of these were exposed to the internet, but CI's
govulncheckwould have failed the next push.
Changed
- New optional variable
USER_DAILY_AUDIO_LIMIT(default1h).0turns
the limit off; any other value must be at leastMAX_MEDIA_DURATION, or the
service refuses to start, because a file longer than the allowance could
never be transcribed. Production does not set it today, so after this
release it runs with a one-hour allowance. - Migration
006_daily_audio_limit.sqladdsjobs.audio_seconds, the
seconds Deepgram billed for the job, and the job statuslimited. A refused
job is neither a failure nor a delivery, so/healthzjobs_failedand the
failure counters do not count it. Rows finished before the migration have no
audio_secondsand count as zero. - A release waits for CI.
release.ymlrunsci.ymlon the tagged commit
and publishes nothing unless it passed. The checks it used to repeat inline
now come from CI itself, so the two cannot drift. When CI cannot run,
release.ymlcan be started by hand with the tag and a reason; that run skips
CI and prints the reason at the top of the release notes. See
docs/releases.md. - The image carries
org.opencontainers.image.versionand.revisionlabels,
sodocker inspectnames the build without starting it. - CI also runs weekly, so a new advisory against a dependency or the standard
library is noticed while nothing is being pushed. - Dependabot opens grouped version updates against
dev.
Container image
ghcr.io/freshlabdev/voicy@sha256:608596d7bd79cf57370773b1c58dd14ac5fa21ce997ddd0b6030b919f82e40cd
v0.0.2
A language picked by hand can be handed back, and every screen is built the same
way as every other screen in the family.
Four panels used to assemble their own markup, the closed statistics tab carried
no mark, and the settings switches were painted with the colour that means
"this is the state you are in" — on the control that turns the state off.
Changed
docs/releases.mdnames the real cause of the403a digest pull used to
answer: the account running the deploy could not read the private package,
not anything about the digest. Pulling the tag first was a workaround for
that, and is no longer required.
Container image
ghcr.io/freshlabdev/voicy@sha256:929f21b9abc7d797b665ace21fa91762423b25bf0162bdaa8a7d66788ce6c39a
v0.0.2-alpha.1
Added
- The language screen offers "Follow Telegram". Picking a language by hand
records a manual choice in the shared Core hub, and manual outranks every
automatic source for ever — so somebody who tapped the wrong flag once read
Voicy in that language until a human edited the database. The button clears
that choice and lets the Telegram client's ownlanguage_codedecide again.
The database role now also needsEXECUTEoncore.clear_language.
Changed
-
Every panel is one shape. Help, Language, Settings and About each built
their own HTML — Help on the argument that a bulleted list is already a shape,
About because the version had to sit next to the name. They all go through the
one panel helper now: bold title, italic one-line hint, substance in a quote.
The About card still leads withVoicy · <version>and the language screen
still says nothing its sixteen buttons already say. -
One colour, one meaning. The open statistics tab was Primary, the colour
that marks the single thing a person came to do; it reports which numbers are
showing, so it is Success now, and Primary is left to the one button that
leads anywhere — Language on the direct-chat home. Both statistics tabs carry
a state glyph, not just the open one, and every switch that is on is Success
like the current language is. Seven coloured switches are loud; a colour that
means one thing on one screen and another on the next is worse.
Container image
ghcr.io/freshlabdev/voicy@sha256:c52b98d02e5b21682a067fb00a4aace0b41f2ebe0cba1ca8f58a6e0e65857946
v0.0.1
Everything that shows or configures now lives behind /start, the panels stopped
wearing one shape that fitted none of them, and a callback can no longer act on
a message it was never shown.
Changed
-
A panel callback only acts on a message it is entitled to act on. Telegram
lets a client send any callback data for any message it can see, not only the
buttons it was shown; the owner id inside the data stopped one person driving
another's panel, but not somebody forging their own id against a public
message. In a group Voicy's public messages are transcripts, soclosecould
have deleted someone else's, and opening a tab could have overwritten one.
Ephemeral messages are unaffected — they are already visible to one person —
and a direct chat is that person's own. -
The direct-chat command menu is one entry:
/start./stats,
/language,/helpand/aboutwere published as commands while already
being buttons on the/startpanel, so every one of them listed the same door
twice. Everything that shows or configures now lives behind/start; a
command is published only when it acts on a message somebody points at, which
is what/vand/vpdo and why the group menu is unchanged. All four still
answer for anyone who types them from memory — they are simply no longer
advertised. -
The panel is two screens instead of one. In a direct chat it keeps every
tab and no longer offersClose: there the whole conversation is the panel,
and closing it deleted the message the reader was looking at. In a group it
offersClose, because there the panel is Voicy's message sitting in somebody
else's feed, and it now shows only the about card and a line explaining/v
and/vp. Settings and the interface language are personal and shared with the
sibling bots through Core, so a group visitor is no longer offered switches
that do not mean in a group what they appear to mean. -
Each panel now carries the shape its content asks for. One "title,
subtitle, quoted body" frame was stretched over five screens of different
kinds, and because the frame required a subtitle, subtitles were invented for
screens that had nothing to put there — "About" was captioned "Voicy". Help is
a list and is no longer boxed in a quote; the language screen is a heading and
one sentence over the picker; settings is a heading and the note the toggles
cannot show; the home card keeps the full frame, which is the one place it was
always right. -
Aboutis the family-standard card. It leads with the product and the
build it is actually running — the same version string/healthzreports — one
line of what Voicy does, thenlabel · valuerows: the recognition engine, the
repository as a link in the text under Apache-2.0, and the admin to write to.
The repository is deliberately a link and not a button: two ways to open one
address are not two actions. Localized in all sixteen languages. -
Button styles mark one thing per screen rather than four scattered
accents. The language tab is highlighted on the direct-chat home, since nothing
else there is readable until the language is right; the language currently in
use is marked in the grid of sixteen; the open statistics tab is marked; and
Closeis destructive. The seven-switch settings grid stays unstyled — its
state is already on the glyphs, and colouring all of it would highlight none of
it. -
One versioning and release document for the whole family.
docs/versioning.md
anddocs/releases.mdare now byte-identical across every Asterfield
repository apart from two clearly marked sections: this repository's own
version line, and the surface where a change here breaks something. They spell
out what each of the three numbers means, what the-alpha.Nsuffix counts,
when alpha becomes beta and when it is legitimate to skip to rc or run a
pre-release in production. -
Pre-releases are now tagged on
dev, notmain. Only stable versions are
tagged onmain, on the merge commit fromdev, somainanswers exactly one
question: what is in production. The test bot runsdev, the production bot
runsmain.release.ymlenforces this and refuses a tag on the wrong branch.
Earlier pre-releases in this repository were tagged onmainunder the
previous rule; they are left as they are.
Removed
- Translation keys that nothing renders any more:
cmd.stats,cmd.language,
cmd.helpandcmd.aboutwith the commands they described,about.title,
about.hintandabout.bodywith the old three-part About panel, and
help.hint,lang.hintandsettings.hint— the subtitles that only restated
their own title or their own body.
Container image
ghcr.io/freshlabdev/voicy@sha256:60bdb59de98f5c82836ba2b021c267bfe3768062c83c6159bf13c3bd7d585f79
v0.0.1-beta.4
Fixed
voicy_telegram_errors_totalno longer counts the startup capability probe.
The probe calls a method with an empty body on purpose and is answered with
a parameter error, so every restart recorded two Telegram errors before the
bot had served anything.
Container image
ghcr.io/freshlabdev/voicy@sha256:28787e90ad9e35785981c8fde084ca6553eb8286279546a1e90ee930e47c1dac
v0.0.1-beta.3
Everything a review of the shared Telegram client turned up, including two
ways the bot token could have escaped.
Security
- Through
tgv0.0.1-alpha.6: the bot token can no longer reach a log line
through an error body echoed by something in front of the Bot API server, nor
through the file path wrapped inside a local-file error. A symlink in the
server's media directory can no longer read a file out of another bot's
directory.
Fixed
- Through the client: an
ok=falseanswer carried on a 2xx is an API error
again, so rate limits and edit-not-modified are classified rather than
slipping through as untyped failures; a startup preflight waits through a
Bot API server that is still booting instead of exiting on its first 502;
andTELEGRAM_READY_WAITnow bounds the whole startup check rather than only
the pauses inside it.
Container image
ghcr.io/freshlabdev/voicy@sha256:339358301cf50104323820bdb4eb1342773a0a381d7e1533fc277351f179c6f5
v0.0.1-beta.2
Voicy's Telegram client is no longer its own, and it will not start against a
server that cannot serve it.
Changed
- Telegram now goes through
github.com/FreshLabDev/tg, the client shared by
the bot family, andinternal/telegramis gone. The transport, retries,
token redaction, rich messages and the 10.3 ephemeral contract were the same
code in three bots and had already drifted apart; they now have one home.
Nothing about Voicy's behaviour changes with them. internal/media.Ofreplacestelegram.Message.Media(). Go cannot put a
method on another package's type, and the order in which attachments are
preferred — speech first — was always a Voicy decision rather than a
Telegram one.voicy_telegram_errors_totalnow also counts transport failures (DNS,
connect, reset). They were invisible before, because only a parsed API error
reached the counter.
Added
- A preflight at startup. Voicy names the methods it cannot work without —
sendRichMessageandeditEphemeralMessageText— and refuses to start when
the Bot API server does not implement them. A server behind the bot answers
404 method not foundto every rich message, which used to mean a bot that
polled happily and answered nothing at all. BOT_API_FILES_DIR, required whenTELEGRAM_API_BASEis self-hosted, and
TELEGRAM_READY_WAIT(default 30s) for a server that is still booting.
Fixed
- Media on a self-hosted Bot API server is read from disk again. Such a server
runs with--local, and a--localserver serves no files over HTTP at all:
its/file/bot<token>/…route answers 404 in every version. Fetching over
the network, introduced in v0.0.1-alpha.10, could therefore never work. Only
this bot's own directory is mounted — the parent holds one directory per bot
named after that bot's token — and the mount is verified at startup instead
of failing on the first voice message.
Operations
- The WS04 stack mounts the bot's own media directory on the Bot API server
and runs as uid 101, the user that server writes as.BOT_API_HOST_DIRand
BOT_API_FILES_DIRname that directory on the host and in the container.
Files are removed after transcription, which a local server never does on
its own. - The mount needs Compose's long volume syntax with a
bindoption. Both
paths contain the bot token, a token contains a colon, and Compose flattens
an option-less long mount back intosource:target:rw, which the daemon
then splits in the wrong places. - Production runs against
telegram-bot-api-next, the FreshLab-built Bot API
10.3 server, while the older shared server keeps the bots that have not
moved. Moving a bot between servers is: stop it,logOuton the server it
leaves, then start it on the new one. A running bot re-registers itself
within a second, so the stop is not optional.
Container image
ghcr.io/freshlabdev/voicy@sha256:40e4f7f6a3a4a7cd7867e24278d649aab8b31b7e95dca5430dce636a938a038e