Skip to content

feat(transcription): offer Voxtral Mini Transcribe 3 and Qwen3 ASR on OpenRouter - #544

Merged
chhoumann merged 95 commits into
masterfrom
feat/transcription-openrouter-models
Nov 16, 2026
Merged

chhoumann merged 95 commits into
masterfrom
feat/transcription-openrouter-models

Conversation

@chhoumann

@chhoumann chhoumann commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

Why

Users should be able to pick from current transcription models without anyone landing on a worse or retired one. This PR adds the two plain OpenRouter models (no speaker labels) that measured best this week against NPR human transcripts.

Summary

The OpenRouter Model row now lists five models, in this order: Grok STT 2.0 (recommended), ElevenLabs Scribe v2, MAI-Transcribe 2, Voxtral Mini Transcribe 3, Qwen3 ASR 1.7B.

Model Mean word error Speaker labels Request Answer Part size
mistralai/voxtral-mini-transcribe-3-0 4.55%, the best measured on OpenRouter no multipart model, file, response_format=verbose_json, no timestamp_granularities (it returns 400 "Provider returned 400") segments[] with start, end, text, no words 12 MB, 900 s (20 MB took 33.5 s)
qwen/qwen3-asr-1.7b 6.05%, about $0.03 per hour no multipart model, file, response_format=verbose_json, timestamp_granularities[]=word timed words[] 12 MB, 900 s (20 MB took 34.5 s)
  • A part that comes back with only segments becomes timestamped paragraphs. Each segment is one cue, timed from the episode's start the same way words are. PartTranscript gains a { segments } variant. The wire shapes stay inside openrouter.ts.
  • An OpenRouter answer with neither words nor segments still fails with "OpenRouter found no speech in this episode." It is not saved as an untimed note. Only OpenAI, whose provider returns text by design, writes untimed notes.
  • If speaker labels are on and one of these models is chosen, PodNotes behaves as it does for any model without labels. The Speaker labels row reads " does not label speakers." and the note is saved without labels, with a notice that stays on screen.
  • One small fix comes along. The notice used to name the provider, so it read "because OpenRouter does not label speakers. These providers do: ..., OpenRouter." It now names the model ("because Voxtral Mini Transcribe 3 does not label speakers."), as the settings row already does. This notice is not on master, so no released wording changes.

Stacked on #542.

Verification

Commands, all with Node 22.23.2 through fnm exec:

  • npm run lint, npm run format:check, npm run typecheck, npm run build, npm run test (124 files, 2,894 tests), npm run test:tooling (48 tests), npm run docs:build: all exit 0.
  • npm run test:e2e -- tests/e2e/journey-resilience.test.ts -t "J-TR-": 11 passed, including the new J-TR-11 (Voxtral) and J-TR-12 (Qwen3 ASR) on Obsidian 1.14.4.
  • npm run size: gzip 107,084 bytes. The base commit 5eebf37, built from a git archive export with the same command, measured 106,873, so this PR adds 211 gzip bytes (728 raw). The gzip budget goes up by exactly those 211 bytes, from 106,970 to 107,181, which leaves the same 97 bytes of headroom as before. scripts/size-baseline.md records this.

Planted defects, each run with npm run plant. Every one failed the named test and passed again once restored:

Plant Caught by
Voxtral sends timestamp_granularities[]=word mistralai/voxtral-mini-transcribe-3-0 gets the upload as its measured multipart request, ... (unit), and J-TR-11 in real Obsidian, where the stub answers it with the measured 400
Qwen3 ASR omits timestamp_granularities (uses Voxtral's request) qwen/qwen3-asr-1.7b gets the upload as its measured multipart request, ...
Part offset dropped for segments (("words" in transcript ? start : 0) + item.start) Voxtral's segments from two parts of at most 900 s become paragraphs timed from the episode's start
Voxtral parts at 1,800 s instead of 900 s the same two-part test
An answer with no words and no segments saved as untimed text an answer with neither words nor segments writes no note
A segments answer treated as untimed (isTimed checks for words only) the Voxtral request test
Qwen3 ASR and Voxtral listed in the wrong order choosing it shows its API key row and its five models, Grok STT 2.0 recommended
Voxtral profile says speakers: true choosing mistralai/voxtral-mini-transcribe-3-0 says Speaker labels are not available, ...
Notice names the provider again (choice.label) with speaker labels on is saved without them, and the notice names the providers that label speakers and stays

Runtime check in this worktree's isolated Obsidian 1.14.4 on Linux (headless, Xvfb), using a fresh vault and a fake key. The .claude/skills/verify/net.mjs fixture provider answered every request, so no real key was involved. The local file was Fixtures/Two Voices.mp3, 70 min of 8 kbps mono, written with the recipe's ffmpeg line (4,200,452 bytes). For each model I chose it in Settings > PodNotes > Model with the keyboard, left Speaker labels on, played the file through Play with PodNotes in the file menu, and ran Transcribe current episode from audio from the palette.

  • Voxtral: five multipart/form-data POSTs of 840,311 to 840,715 bytes (840 s each), one after the other, and no reel. The fixture answers a word-times request with a 400, and the note was saved, so the request did not ask for word times. The note has paragraphs at 00:00:00, 00:00:12, 00:14:00 (time=840.096), 00:14:12, 00:28:00 and so on up to 00:56:12, with no speaker labels. Clicking 00:28:12 moved the player to 1,693.66 s, 1.5 s past the link's 1,692.192 while it kept playing. The notice read "Saved without speaker labels, because Voxtral Mini Transcribe 3 does not label speakers. These providers do: ElevenLabs, Deepgram, OpenRouter." dev:errors: none.
  • Qwen3 ASR: five POSTs of 840,409 to 840,813 bytes. Each is 98 bytes larger than Voxtral's, the size of its timestamp_granularities[] field. The note times match Voxtral's, and the notice names Qwen3 ASR 1.7B. data.json stored transcript.model: "qwen/qwen3-asr-1.7b".
Settings, Voxtral Note, Voxtral
Settings with Voxtral Mini Transcribe 3 chosen Voxtral note with episode-absolute timestamps
Settings, Qwen3 ASR Note, Qwen3 ASR
Settings with Qwen3 ASR 1.7B chosen Qwen3 ASR note with episode-absolute timestamps

Two things in the screenshots come from elsewhere. The warning icon on the key is Obsidian's own "Secrets are not encrypted", because headless Linux has no keyring. The doubled "Two Voices" is Obsidian's inline title shown above the template's # {{title}}.

I have not run this against the real OpenRouter API. The lead verifies with a real key.

Real OpenRouter key (lead, after review)

Obsidian 1.14.4, Linux, palette command, speaker labels off, a 20 min local clip of E231 (expect 2 parts at 900 s), a pass-through request log:

  • Voxtral Mini Transcribe 3: one catalog GET, two POSTs of 8.10 and 6.94 MB, saved in 25 s. 3,333 words, segment times running 00:00:19 to 00:19:42 across both parts.
  • Qwen3 ASR 1.7B before the fix below: the same two POSTs, saved in 22 s, but the note had no punctuation at all ("Hey everyone and welcome to another episode of The Five These Times My name is ...") and paragraphs cut every 60 s. Its words carry no punctuation (0 sentence ends in 3,785 words on Planet Money), while its text does (322).
  • Qwen3 ASR after the fix: saved in 33 s, "Hey everyone, and welcome to another episode of The Five These Times. My name is ...", 340 sentence ends in 3,366 words, paragraphs at 00:00:19, 00:00:54, 00:01:24. dev:errors none.

Review fix (lead)

  • Qwen3 ASR words take their punctuated token from text, matched in order with three tokens of lookahead, only when no word carries punctuation. Plants: keep the bare word on a match; one token of lookahead. Each is caught by Qwen's words, which it lists without punctuation, take their punctuation from its text.
  • Gzip budget raised by the measured 100 bytes to 107,281 (107,184 measured).
  • Checks: lint, format, typecheck, build, size, 2,895 unit tests, 48 tooling tests.

Release impact

There are no settings or storage changes. transcript.model can now hold two more OpenRouter IDs. An older PodNotes that reads them treats them as unknown and runs its recommended model, which is existing behavior. The speaker-label notice now names the model instead of the provider, and that notice has not been released yet.

Note

Add Voxtral Mini Transcribe 3 and Qwen3 ASR 1.7B to OpenRouter transcription

  • Registers both models as multipart-only OpenRouter providers with 12,000,000-byte and 900-second part limits. Neither model produces speaker labels. See openrouter.ts
  • Qwen3 ASR requests word timestamps and restores punctuation from the response text via a new token-matching helper in openrouter.ts
  • Voxtral requests segment times only. cues.ts now builds timed cues from segment responses as well as word responses, and types.ts adds a Segment interface
  • Raises bundle budgets in size.mjs to 330,546 raw and 108,180 gzip bytes for the added model code
  • Behavioral Change: the speaker-label warning now names the selected model (e.g. GPT Transcribe) instead of the provider label in transcribe.ts; malformed segment entries are discarded by segmentOf

Macroscope summarized c69d365.


Note

Medium Risk
Touches transcription provider requests, response parsing, and note formatting for new OpenRouter models; no auth or storage schema changes, but incorrect parsing or part limits could affect saved transcripts.

Overview
Adds Voxtral Mini Transcribe 3 and Qwen3 ASR 1.7B to the OpenRouter transcription model list, each with smaller upload caps (12 MB / 900 s), no speaker labels, and multipart-only audio (no URL transcription).

OpenRouter provider gains model-specific request shapes (Voxtral: segment times only—word-time requests fail; Qwen: word times) and parsing for segment-based answers plus Qwen punctuation restoration from response text onto bare words. Transcript assembly (cues.ts, types) now builds timed notes from segments as well as words, with the same episode-offset timing across parts.

When speaker labels are enabled but the model cannot diarize, the completion notice names the model (e.g. GPT Transcribe) instead of the provider. Bundle size budgets rise to match the added logic (~546 raw / ~781 gzip bytes). E2E fixtures, verify docs, and broad unit/journey tests cover the new models and Qwen edge cases.

Reviewed by Cursor Bugbot for commit c69d365. Bugbot is set up for automated code reviews on this repo. Configure here.

gpt-4o-transcribe-diarize rejects requests with more than 1400 s of audio,
but MP3 chunks were split by size alone. Any MP3 under about 120 kbps put
more than 1400 s in a 20 MB chunk, so every chunk failed.

For diarization, split MP3 at frame boundaries into the fewest equal parts
that stay under both 20 MB and 1400 s. Whisper keeps its size-only chunks.

Fixes #532
A chunk that OpenAI rejected with a status such as 400 or 401 was sent
three times, although each send fails the same way. A long diarized
episode went up three times before the error showed.

Retry only the statuses openai-node retries: 408, 409, 429 and 5xx, plus
network errors.

Refs #532
gpt-4o-transcribe-diarize takes up to 0.58 s per second of audio, measured
over 81 requests, so the 1,072 s parts a 54 min episode got would take about
10 min each, at the request timeout. Parts are now at most 600 s, which the
model answers in under 6 min.

Refs #532
Each test passes on the hand-written code. Together they catch wrong
derivations that no existing test sees: the -1 SecretStorage ID suffix
that released versions accept, the provider an export error names, the
kind a settings key row adopts a secret for, and the key migration
notices.
CREDENTIALS in src/types/Credentials.ts holds each kind's settings
field, export key and label. The credential repository, settings
decoder, settings transfer, settings tab and key migration notice loop
over it. IPodNotesSettings takes its key fields from it, so a kind with
no default in DEFAULT_SETTINGS fails typecheck. The plaintext key map
from before 2.19.0 stays fixed.
The published transcript renderer is the reader of the format that stays
when the diarization folder goes. The default template now lives only in
DEFAULT_SETTINGS, so src/constants.ts imports no service code.
src/services/transcription/ turns an episode into a transcript with any
provider. A provider is one file: its models, its request, and how to read
its answers. OpenAI (gpt-transcribe) and Deepgram (nova-3) are the first two.

The engine owns what providers share:

- audioSource.ts decides whether a provider may be given the episode's
  address. Only an address that a saved public feed lists, that carries no
  credential, and that names a public host passes. All other audio is
  uploaded.
- split.ts is audioChunker.ts with the caps as arguments and a start time
  on each part.
- http.ts sends a request at most three times, and only after a 429, a 5xx,
  or a lost connection. Its failure sentences never include the provider's
  own text, which can quote a private address.
- cues.ts turns words or text into the shapes of a published transcript.

Nothing calls the engine yet. Lint rejects an import of the folder's files
from outside it, except its index.
TranscriptionService keeps its queue, the published-transcript path, the
progress notice, and note saving. It calls the engine for audio.

- OpenAI transcribes with gpt-transcribe. whisper-1 and
  gpt-4o-transcribe-diarize, which OpenAI shuts down on 2027-02-26, are
  gone. A tooling test searches the built main.js for them.
- With speaker labels on and OpenAI selected, the transcript is saved
  without labels. A notice that stays until clicked names the providers
  that label speakers.
- Deepgram transcripts go through renderTranscriptBody, so each paragraph
  has a timestamp. Speakers are numbered in the order they first speak.
- Deepgram gets the address of a public episode and downloads it itself.
- A failure notice stays until clicked. Parts go out one at a time, and a
  401, 402, 404, or 400 stops the job before the next part.

Deletes src/services/diarization/, openaiTranscription.ts, the Whisper
worker pool, and createBinaryChunkFiles. getEpisodeAudioBuffer no longer
returns a basename, because uploads are named for their format only.
net.mjs answers OpenAI with { text } and Deepgram with the words of two
speakers, and logs the content type of each request. fixtures.mjs serves a
feed at priv3.supercast.com, an address that PodNotes keeps private. The
transcripts recipe covers OpenAI in parts, and Deepgram by address and by
upload.
The marketplace review test called TranscriptionService.waitForRetry, which
the engine replaced. An aborted run now rejects through waitForLifecycle, so
the test aborts a pending wait there and expects the same Error.
An aborted job no longer downloads the episode or decodes it into parts.
Until now only its requests were refused.

A 429 or a 5xx that asks to wait over a minute is sent once. A sooner send
would only upload the audio to be refused again.
A provider's name was written twice: once in its credential row and once in
its provider file. The row now owns it, so a provider file holds only its
models and its request. Each model gains the name the settings will show.
The provider was derived from the speaker settings: Deepgram with speaker
labels on, OpenAI otherwise. It is now stored as transcript.provider, with
transcript.model beside it, where "" follows the provider's recommended model.
schemaVersion stays 2 and both fields are additions, so a device on 2.24.0
keeps loading the library and keeps the fields on its own saves.

A library that has no transcript.provider gets the provider it ran with:
Deepgram for speaker labels with Deepgram, with or without its key, and OpenAI
otherwise. The migration reads only the stored document, adds no decode
warning, and gives the same answer on every load until a save stores it.

transcript.diarization.provider belongs to versions up to 2.24.0. This version
reads it for that migration alone, never writes it, keeps it in data.json as
found, and leaves it out of settings exports. Importing a 2.24.0 export that
labels speakers with Deepgram selects Deepgram.

A provider or model ID this version does not have is kept as stored and shown
under that name. An unknown provider refuses to transcribe with a notice. An
unknown model runs the recommended model and the notice says so.

The Transcripts settings now read: Transcription provider, the chosen
provider's API key, Model, Speaker labels, Speaker label format, then the path
and template. Speaker labels is disabled for a model that cannot label
speakers. The Diarization provider row is gone.
ElevenLabs is a credential row, one provider file and a registry line. Its key
reference is the new root field elevenLabsSecretId, stored in SecretStorage as
podnotes-elevenlabs-api-key, and a settings export carries the key as
secrets.elevenLabs. The export envelope stays at version 2.

A public episode goes to ElevenLabs by its address, so nothing is downloaded.
Any other episode is uploaded in one request. Word times and speakers reach
the note, and turning speaker labels off asks for none.

ElevenLabs answers 401 for a used-up quota as for a rejected key, and 400 for
a model the key may not use as for an address it could not download, so the
provider reads detail.code. Only the last of these is followed by an upload.
A request waits 10 minutes for its answer.

A new install starts on ElevenLabs with speaker labels on. A library saved up
to 2.24.0 that holds no OpenAI key and has speaker labels off moves to
ElevenLabs too, because it never set transcription up. A library that never
stored the speaker labels choice keeps them off.

The notice for a model without speaker labels now names ElevenLabs first.
A journey now seeds its transcription provider with its key, because a new
vault is on ElevenLabs. The OpenAI journeys choose OpenAI. A new journey runs
the default: a public episode goes to ElevenLabs by its address, no audio is
fetched, and the note has timestamps and speaker labels. The ElevenLabs stub
answers with the words ElevenLabs returns and records what each request named.

The settings tab test walks the new Transcripts rows, and the key recovery
test links the ElevenLabs key, the one a new vault shows.
The fixture provider answers api.elevenlabs.io with ElevenLabs' words, two
speakers with the second 12 s in, and logs the address a request named in
place of a file. The transcript guide follows the new settings rows and adds
the ElevenLabs runs and a 2.24.0 library check.
…choice

The provider row decided for itself whether the stored provider is one this
version has, and the Speaker label format row read the stored speaker setting.
Both now ask transcriptionChoice, as the key rows, the model row and the
Speaker labels row do. The engine no longer exports a type nothing imports.
The Transcripts tests looked each control up by a CSS selector inside its row
and asserted on a hidden row's inline style. They now query by role and by the
row's name, as a user or a screen reader finds a control, so a hidden row is
one the query does not find.
…ings measure

The transcription engine left main.js at 103,795 gzip bytes. The provider and
model settings, their migration and the ElevenLabs provider add 866, which
takes it to 104,661 and over the 104,000 budget. The budget rises by those 866
bytes to 104,866, so the 205 bytes of headroom stay as they were.
…m library

2.24.0 imports the diarization object whole and fills a missing provider
with OpenAI, so omitting the field moved Deepgram libraries to OpenAI.
Imports here keep the provider this library stored.
A library migrated with OpenAI speaker labels on could not turn them off,
so every transcription repeated the notice.
… one

The tests for a provider and a model a newer PodNotes stored used OpenRouter
and Grok STT 2.0, which this version is about to run. They now use AssemblyAI
and universal-3-6-pro.
OpenRouter is a credential row, one provider file and a registry line. Its
key reference is the new root field openRouterSecretId, stored in
SecretStorage as podnotes-openrouter-api-key, and a settings export carries
the key as secrets.openRouter.

It offers three models, each with the request measured to give word times
and speakers:

- Grok STT 2.0, recommended: base64 JSON with verbose_json and word and
  segment times, routed to xAI with provider.only, which takes diarize from
  provider.options.xai. A request routed elsewhere is answered with no
  speakers.
- ElevenLabs Scribe v2: multipart with diarize beside the model, or JSON with
  input_audio.url for an episode a saved public feed lists.
- MAI-Transcribe 2: multipart with diarize beside the model.

With speaker labels off, no request carries diarize. Sounds Scribe v2 lists
as audio_event items are left out.

OpenRouter takes 25 MB a request and gives the model 60 s, so an upload goes
in parts of at most 18 MB and 1,800 s. Parts of 24 MB were answered in 20 to
31 s.

A 402 says the account is out of credit, a 401 that the key was rejected,
and a 404 that the model is not offered. A 404 whose message names the
account's data policy or zero data retention says the model is not offered
under the account's OpenRouter privacy settings. Each ends the job at the
first part, and none quotes OpenRouter.
…t each part

A note saved with speaker labels asked for and none returned now says so:
"Saved without speaker labels, because Grok STT 2.0 returned none." OpenRouter
answers 200 with no speakers when a request reaches a route that ignores the
speaker option, so this is checked for every provider.

Labels from two parts already name different speakers. An episode sent in
parts now also says that its labels restart in each part, so one person can
have several labels, until labels are merged across parts.
…ll lists

PodNotes reads OpenRouter's public catalog of transcription models, without
the key, when the Model row draws and when an OpenRouter job starts. It keeps
the read in memory for 24 hours. A read that fails is not kept, so the next
one asks again, and the tested list is offered as it is.

The catalog only takes models away. A model it does not list is not offered,
and one whose expiration_date has come is skipped. A model with a later end
date reads "Grok STT 2.0, ends <date>" in the Model row. A catalog that would
leave no model is not followed.

A chosen model the catalog no longer lists stays selected in the Model row as
"<label>, no longer offered". A job runs the first offered model instead and
says "OpenRouter no longer offers ElevenLabs Scribe v2, so PodNotes used Grok
STT 2.0." The row redraws when a read lands while it is open.

A provider gets its catalog through an optional catalog() beside failure(),
so the registry owns the cache and the filter for any provider that has one.
A stub answers OpenRouter's transcriptions with two speakers, the second 12 s
in, and its model catalog with the three models PodNotes offers. It records
what each request named and a hash of the audio it carried, decoded from
base64 for JSON.

J-TR-09 plays a public episode on Scribe v2: one JSON request names the
episode's address, and the episode's audio is never fetched. J-TR-10 plays it
on the recommended Grok STT 2.0: one JSON request carries the audio as base64,
the same bytes the fixture server holds. Both notes have timestamps and
numbered speakers, and the catalog is read once, with no Authorization.

A journey seeds any provider by its credential row, and a model with it.
…eads them

The fixture provider answers openrouter.ai transcriptions with verbose_json
words from two speakers, the second 12 s in, and the model catalog with the
three models PodNotes offers. The log keeps openrouter.ai paths and reads the
address a JSON request names in input_audio.url.

The transcript guide adds the OpenRouter rows, its runs (Scribe v2 by
address, Grok by base64, a long MP3 in parts) and the notices that now name
OpenRouter.
@chhoumann

Copy link
Copy Markdown
Owner Author

FAIL
Verified head d8099b5 by an independent verifier (GPT 6 Astra).

  • P2, duplicated punctuation in the shared reader. With text: "Hello… Next!" and timed words Hello…🙂 and Next at 1–1.15 and 1.2–1.35 seconds, the saved note reads Hello…🙂… Next!. Master preserves one ellipsis. The whole-answer guard recognizes only ASCII marks, while the check at line 301 misses punctuation before emoji.
  • Reproduced through the palette with Qwen, Grok, Scribe v2 and MAI, and through the direct command with Qwen. Chinese 你好。🙂 becomes 你好。🙂。; 你好!!🙂 becomes 你好!!🙂!!. Dashes and opening quotes beside emoji also double. These existing-model regressions are absent on master.
  • Compared master f457c5b0 against the pinned head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head matrix cases through Settings mouse/keyboard selection and palette transcription, with labels on/off, public/private fixture feeds on port 28368, a 70-minute MP3 and a 40-minute M4A. Only fake keys and fixture responses were used.
  • Independently decoded 302 uploads within their model caps. New-model MP3 jobs use five parts; M4A jobs use eighteen. Voxtral omits word-time requests; Qwen requests them. Part offsets, omitted labels and model-specific notices pass. The 24 ordinary Grok/Scribe/MAI head notes are byte-identical to master.
  • All 141 replay notes preserve original word characters and add only punctuation after Markdown unescaping. The earlier symbol-loss, split-compound/repetition and separate-punctuation failures pass. Both cached Qwen answers preserve all 2,262/3,785 words and restore 301/311 and 612/616 Latin marks. Chinese/Japanese mark runs, mixed scripts, contractions, numbers and quotes were exercised.
  • Both 17,500-word notes preserve every word; single fixture runs took 1.43 and 2.61 seconds including driver overhead. The accepted 210-token omission limit remains. Withdrawing either new model marks it unavailable, falls back to Grok with a notice and hides it after another selection. Untimed answers save no note.
  • All requested checks pass through fnm exec --using 22.23.2: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,942 tests), and npm run test:tooling (56 tests). The installed Node entry reports v22.23.4. Bundle size is exactly 330,529 raw / 108,136 gzip bytes.
  • dev:errors reports none. No tracked files changed. Copied drivers, responses, saved notes, decoded-upload evidence and inspected screenshots are in this worktree's .obsidian-e2e-artifacts/review544/, starting with verification.md and duplicate-proof.json. The fixture server and isolated Obsidian are stopped; the stop dry run reports pids: none.

…igit

A mark hidden behind an emoji, as Hello…🙂 or 你好。🙂, slipped past the
already-punctuated checks, and the word gained it twice. It also reached Grok,
Scribe v2 and MAI, which share the reader. A word now gains an opening mark
only if it starts with a letter or digit, and closing punctuation only if it
ends with one, so a word with any mark, symbol or emoji at that end keeps it
as given.
@chhoumann

Copy link
Copy Markdown
Owner Author

On the d8099b5 verdict: fixed in c2b1c7c with one rule instead of more mark sets. A word gains an opening mark only if it starts with a letter or digit, and closing punctuation only if it ends with one. A word that has any mark, symbol or emoji at that end keeps it as given, so no mark can be added twice, whatever script or emoji is beside it. That covers Qwen, and Grok, Scribe v2 and MAI too, since all four use this reader.

  • New test "adds no second mark to a Qwen word that has one beside an emoji". It covers “🙂Hi, Hello…🙂 and 你好。🙂, and was red on d8099b5.
  • Plants caught: the opening-edge check removed, and the old closing check. Each is caught by this test.
  • Cost: a word that ends in a symbol, such as 5% at the end of a sentence, gains no period. The real Qwen3 words carry no symbols. Both answers are unchanged: 301 of 311 and 612 of 616 marks, every word kept.
  • test (2,943) and test:tooling (56) pass. Budgets are 330,557 raw and 108,141 gzip, measured.

chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader adds marks only where a word ends in a letter
or digit, as #544 now does.
@chhoumann

Copy link
Copy Markdown
Owner Author

FAIL
Verified head c2b1c7c by an independent verifier (GPT 6 Astra).

  • Reproduced a doubled opening quote at openrouter.ts:301. With text: '"1️⃣Hello world" Next!' and timed words ['1️⃣"Hello', 'world', 'Next'], the saved note reads "1️⃣"Hello world" Next!. The first word gains a second opening quote. Starts are 1, 1.2 and 1.4 seconds, each lasting 0.15 seconds.
  • The information emoji reproduces it too: ℹ️"Hello becomes "ℹ️"Hello. Both emoji begin with a Unicode letter or digit, so the new first-codepoint guard accepts them. Reproduced through the palette on Qwen, Grok, Scribe v2 and MAI, and through the direct audio command on Qwen. Master preserves one opening quote on all three existing models.
  • Compared master f457c5b0 with this head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head matrix cases through Settings mouse/keyboard selection and palette transcription. Sources were public/private fixture feeds on port 29810, a 70-minute MP3 and a 40-minute M4A, with labels on/off. Only fake keys and fixture responses were used.
  • All 302 captured uploads decode within their model caps. New-model jobs use five MP3 parts and eighteen M4A parts. Voxtral omits word-time requests; Qwen requests them. Part offsets, omitted labels and model-specific notices pass. All 24 ordinary Grok/Scribe/MAI head notes are byte-identical to master.
  • All 161 replay notes preserve supplied word characters and add only punctuation after Markdown unescaping. Earlier symbol-loss, split-compound, separate-punctuation and ordinary emoji failures are fixed. Cached Qwen answers retain all 2,262/3,785 words and 301/311 and 612/616 Latin marks. Both 17,500-word notes retain every word; single fixture runs took 2.48/2.49 seconds including driver overhead. The accepted 210-token omission and terminal 5% limitations remain.
  • Withdrawing either model marks it unavailable, falls back to Grok with a notice and hides it after another selection. Untimed answers save no note. Checked saved notes, persisted settings, notices and screenshots. Every captured dev:errors result reports none.
  • All requested checks pass through fnm exec --using 22.23.2: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,943 tests), and npm run test:tooling (56 tests). That installed Node entry reports v22.23.4. Bundle size is 330,557 raw / 108,141 gzip bytes.
  • No tracked files changed. Copied drivers, responses, notes and inspected screenshots are under /home/christian/.t3/worktrees/PodNotes/c33e0a5e-40bb6b9579afda51/.obsidian-e2e-artifacts/review544/, starting with verification.md, duplicate-proof.json and note-screenshots.png. The fixture server and isolated Obsidian are stopped; the stop dry run reports pids: none.

Emoji such as 1️⃣ and ℹ️ start with a digit or letter, so a word like 1️⃣"Hello
passed the edge check and gained a second quote, with Grok, Scribe v2 and MAI
too. Only a word of letters and digits, with apostrophes and hyphens between
them, now gains marks; any other word keeps its text as given.
@chhoumann

Copy link
Copy Markdown
Owner Author

On the c2b1c7c verdict: fixed in 360cccb. Checking a word's ends kept losing to emoji that begin with a letter or digit, so the rule is now about the whole word. Only a plain word, made of letters and digits with apostrophes or hyphens between them, gains marks. Any word with a mark, a symbol or an emoji anywhere keeps its text as given. No mark can be added beside an emoji, keycap or not, for Qwen, Grok, Scribe v2 or MAI.

  • The emoji test now also covers 1️⃣"Go and ℹ️"Info, and was red on c2b1c7c. Plants caught: the plain check dropped from the opening edge, and the plain pattern loosened back to "starts with a letter or digit".
  • Cost: unchanged from the previous round. A word with a symbol, such as C++ or 5%, gains no mark. The real Qwen3 words are all plain, and both answers are unchanged: 301 of 311 and 612 of 616 marks, every word kept.
  • test (2,943) and test:tooling (56) pass. Budgets are 330,566 raw and 108,168 gzip, measured.

chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader adds marks only to plain words, as #544 now does.

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit 360cccb. Configure here.

Plain words took letters and digits only, so 3.5, $5, 3:30, 5% and 1,000 kept
no comma or period from the text. A word now may also hold a number's .,:/
between digits, a currency sign before and % after, none of which is a mark
it can gain twice.
chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader gives numbers, amounts, times and percentages
their punctuation, as #544 now does.
@chhoumann

Copy link
Copy Markdown
Owner Author

FAIL
Verified head dea6eba by an independent verifier (GPT 6 Astra).

  • P2 defect in openrouter.ts:306. With text: "Hello, world!" and timed words ["Hello", ",world"], head saves Hello, ,world; master saves Hello ,world. Starts are 1 and 1.2 seconds, each lasting 0.15 seconds. The per-word guard misses punctuation already attached to the neighboring word.
  • Reproduced through the palette on Qwen, Grok, Scribe v2 and MAI, and through Qwen's direct audio command. Hello …Next becomes Hello… …Next; 你好 。再见 becomes 你好。 。再见. Opening quotes also double across a boundary. All 24 targeted head cases reproduce; master preserves each original mark once.
  • Compared master f457c5b0 with the pinned head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head source cases through Settings mouse/keyboard, file menus and palette commands. Sources were public/private fixture feeds on port 26086, a 70-minute MP3 and a 40-minute M4A, with labels on/off. Fake keys only.
  • All 302 captured uploads decode within their model caps. New models use five MP3 parts and eighteen M4A parts; Voxtral omits word-time requests and Qwen requests them. Part offsets, omitted labels and model-specific notices pass. The 24 ordinary existing-model head notes are byte-identical to master.
  • Replayed all 58 distinct saved inputs from earlier verifiers, plus new boundary and numeric cases. All 305 replay notes preserve original word characters and add only punctuation after Markdown unescaping. Cached Qwen answers retain all 2,262/3,785 words and restore 301/311 and 612/616 ASCII marks. Earlier symbol-loss, separate-punctuation and emoji failures are fixed.
  • Both 17,500-word notes retain every word; single fixture runs took 2.45/2.54 seconds including driver overhead. Withdrawing either new model marks it unavailable, falls back to Grok with a notice and hides it after another selection. Untimed answers save no note.
  • All requested checks pass through fnm exec --using 22.23.2: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,944 tests), and npm run test:tooling (56 tests). The installed Node entry reports v22.23.4. Bundle size is 330,583 raw / 108,196 gzip bytes, exactly at both budgets.
  • Every captured dev:errors result reports none. No tracked files changed. The fixture server and isolated Obsidian are stopped; the dry run reports pids: none. Copied drivers, responses, notes and inspected screenshots are under /home/christian/.t3/worktrees/PodNotes/45f1ef8d-f3921b5478aabb94/.obsidian-e2e-artifacts/review544/, starting with verification.md, boundary-runtime.mjs and duplicate-proof.json.

…r end

A model that attached a mark to the start of the next word, as Hello ,world,
slipped past the check for punctuated answers, and the plain word before it took
the same mark from the text. Any word with a mark before its first letter or
digit or after its last now leaves the whole answer as given. Apostrophes do
not count, so dogs' still gets its sentence's punctuation. The opening and
closing sets are now written once.
@chhoumann

Copy link
Copy Markdown
Owner Author

On the dea6eba verdict: fixed in 474c515. A model that attaches a mark to either end of a word punctuates its own words, so the check for an already punctuated answer now looks at both ends of every word. Any word with a quote, bracket, dash or sentence mark before its first letter or digit or after its last leaves the whole answer as given. That covers ,world, …Next, 。再见 and Say" on Qwen, Grok, Scribe v2 and MAI. Apostrophes do not count, so a possessive dogs’ still lets the answer be punctuated. The opening and closing sets are written once and shared by the matcher and the check.

  • New tests:
    • "adds nothing to Qwen's words when one starts with a mark the word before could take" (your four boundary shapes) was red on dea6eba.
    • "restores punctuation to Qwen's words when one of them ends in an apostrophe"
  • Plants caught: the old end-only check, and curly apostrophes counted as marks. The emoji test now covers the keycap and ℹ️ cases that only the plain-word rule guards. The other emoji shapes are covered by the new check.
  • The real Qwen3 answers are unchanged: 301 of 311 and 612 of 616 marks, every word kept. test (2,946) and test:tooling (56) pass. Budgets are 330,723 raw and 108,274 gzip, measured.

chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader restores nothing when any word has a mark at
either end, as #544 now does.
@chhoumann

Copy link
Copy Markdown
Owner Author

FAIL
Verified head 474c515 by an independent verifier (GPT 6 Astra).

  • P2, duplicated punctuation at openrouter.ts:242. With text: "Hello, 1️⃣world!" and words ["Hello", "1️⃣,world"], head saves Hello, 1️⃣,world; master saves Hello 1️⃣,world. Starts are 1 and 1.2 seconds, each lasting 0.15 seconds. The emoji's digit hides the existing comma from the guard, and the neighboring plain word gains another.
  • ℹ️,world, keycap ellipses/CJK periods, and opening quotes reproduce it. Single quotes also double: text: "Say 'Hello world'", words ["Say'", "Hello", "world"] save Say' 'Hello world'; curly quotes behave likewise. Seven shapes fail on each of Qwen, Grok, Scribe v2 and MAI through the palette. Qwen's direct audio command reproduces both emoji commas and the single quote. Master preserves each original mark once.
  • Compared master f457c5b0 with this head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head source cases through Settings mouse/keyboard, file menus and palette commands. Sources were public/private fixture feeds on port 29023, a 70-minute MP3 and a 40-minute M4A, with labels on/off. Fake keys only.
  • All 302 captured uploads decode within their model caps. New models use five MP3 parts and eighteen M4A parts, with correct timestamp-request options, part offsets, omitted labels and model-specific notices. All 24 ordinary existing-model notes are byte-identical to master. Both catalog-withdrawal cases pass; untimed answers save no note.
  • Replayed all 68 distinct prior saved inputs. All 404 replay notes preserve original word characters and add only punctuation after Markdown unescaping. Cached Qwen answers retain all 2,262/3,785 words and restore 301/311 and 612/616 Latin marks; only timestamps were scaled. Earlier plain-boundary, symbol-loss, split-compound, separate-punctuation and emoji cases pass.
  • Both 17,500-word notes retain every word; single fixture runs took 2.64/2.51 seconds including driver overhead. The accepted 210-token omission limit remains. Curly contractions can take a nearby filler comma (Well, we’ll see. becomes we’ll, see.), classified as a nonblocking note under the requested criteria.
  • All requested checks pass through fnm exec --using 22.23.2: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,946 tests), and npm run test:tooling (56 tests). The installed Node entry reports v22.23.4. Bundle size is 330,723 raw / 108,274 gzip bytes.
  • Every captured dev:errors report is clear. No tracked files changed. The fixture server and isolated Obsidian are stopped; the dry run reports pids: none. Evidence is under /home/christian/.t3/worktrees/PodNotes/acbf0775-5bb543cc6bc1af17/.obsidian-e2e-artifacts/review544/, starting with verification.md, novel-runtime.mjs, novel-comparison.json and direct-boundary-keycap-note.md.

Checking each word's edges kept missing shapes that hid a mark from the check,
such as a keycap emoji before a comma or a single quote on the word before, and
a plain neighbor then took the same mark from the text. Marks now come from the
text only when every word is plain: letters and digits, with apostrophes,
hyphens or a number's .,:/ between them. Then no word has a mark of its own,
and each token's marks go to one word, so none can be added twice. An answer
with any other word is kept as the model gave it.
@chhoumann

Copy link
Copy Markdown
Owner Author

On the 474c515 verdict: fixed in 648738c by changing the rule rather than patching another shape. Marks now come from the text only when every word is plain: letters and digits, with apostrophes, hyphens or a number's .,:/ between them, and an optional currency sign or %. In such an answer no word has a mark of its own. Each token's opening mark goes to the one word that starts it and its closing marks to the one word that ends it, so no mark can be added twice, across a boundary or not. If any word carries anything else (an emoji, a symbol, a quote, an end apostrophe), the model punctuates its own words and the answer is kept as given. That is what master does for Grok, Scribe v2 and MAI.

  • The real Qwen3 answers are all plain (6,047 words, none with an emoji, symbol or end apostrophe) and unchanged: 301 of 311 and 612 of 616 marks, every word kept.
  • New test "keeps every Qwen word as given when one has a symbol, an emoji, a quote or an end apostrophe" covers C++, 1️⃣,world, Say' and dogs’. It replaces the emoji and apostrophe tests, whose partial restoring no longer applies. The numbers test gained the -3 against 3 case.
  • Plants caught: the all-plain check removed, the plain pattern without numbers, and a plain pattern that only checks the first character.
  • The edge-mark check and the per-word plain checks are gone, so the bundle shrank. Budgets go down to the measured 330,501 raw and 108,158 gzip. test (2,944) and test:tooling (56) pass.
  • Accepted: a part in which Qwen writes a possessive like dogs' or any symbol gets no punctuation from the text.

chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader restores punctuation only when no word has any,
as #544 now does.
@chhoumann

Copy link
Copy Markdown
Owner Author

FAIL
Verified head 648738c by an independent verifier (GPT 6 Astra).

  • P2, duplicated punctuation at openrouter.ts:233. With text: "Hello, world again!" and words ["Hello", "world,again"], master saves Hello world,again; head saves Hello, world,again. Start/end times are 1/1.15 and 1.2/1.35 seconds. Each input representation has one comma; head has two. PLAIN_WORD accepts .,:/ between letters, bypassing the whole-answer guard despite the stated exception being for numbers.
  • Reproduced through the palette on Qwen, Grok, Scribe v2 and MAI, and through Qwen's direct audio command. Within-word duplication also occurs: text: "Helloworld, Next!", words ["Hello,world", "Next"] saves Hello,world, Next!. Master preserves the single comma. Period and colon cases reproduce too.
  • Compared master f457c5b0 and this head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head source cases through Settings mouse/keyboard, file menus and palette commands. Sources were public/private fixture feeds on port 36051, a 70-minute MP3 and a 40-minute M4A, with labels on/off. Fake keys only.
  • Independently decoded all 302 uploads within their model caps. New models use five MP3 parts and eighteen M4A parts, with correct timestamp-request options, offsets, omitted labels and model-specific notices. All 24 ordinary existing-model notes are byte-identical to master.
  • Replayed all 76 distinct prior saved inputs. All 355 replay notes preserve original word characters and add only punctuation after Markdown unescaping. Earlier emoji, quote and boundary failures pass. Cached real Qwen answers retain all 2,262/3,785 words and restore 301/311 and 612/616 marks; only timestamps were scaled.
  • Both 17,500-word notes retain every word; single fixture runs took 2.58/1.58 seconds including driver overhead. The accepted 210-token omission limit remains. Withdrawing either new model marks it unavailable, falls back to Grok with a warning and hides it after another selection. Untimed answers save no note.
  • All requested checks pass through fnm exec --using 22.23.2: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,944 tests), and npm run test:tooling (56 tests). The installed Node entry reports v22.23.4. Bundle size is exactly 330,501 raw / 108,158 gzip bytes.
  • Every captured dev:errors report is clear. No tracked files changed. Inspected screenshots, notes, requests and persisted settings. The fixture server and this worktree's Obsidian are stopped; the dry run reports pids: none. Copied drivers and evidence remain in review worktree ff298086-bbd7bddc6a050e97.

…igits

The plain pattern let a comma, period, colon or slash sit between letters, so
world,again counted as plain, kept its comma, and the word before it took the
same comma from the text. Those separators now belong to numbers alone.
@chhoumann

Copy link
Copy Markdown
Owner Author

On the 648738c verdict: fixed in c245255. The plain pattern now has two shapes: letters and digits with apostrophes or hyphens between them, or a number with .,:/ between its digits, an optional currency sign before and % after. world,again, Hello,world and Time:now are no longer plain, so an answer that holds one is kept as given. 3.5, 1,000, 3:30, $5 and 5% still take punctuation. New test "keeps every Qwen word as given when one has a comma or colon between letters" was red on 648738c, and the old pattern is caught. The real Qwen3 answers are unchanged: 301 of 311 and 612 of 616 marks, every word kept. test (2,945) passes. Budgets are 330,525 raw and 108,164 gzip, measured.

chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader counts .,:/ as part of a plain word only between
digits, as #544 now does.
A word with a combining mark, as a decomposed é or any Hindi word, failed the
plain pattern, so its whole answer got no punctuation from the text. Letters
now include combining marks, in the pattern and the matching keys, and the
Devanagari danda joins the closing marks.
chhoumann added a commit that referenced this pull request Nov 16, 2026
The shared verbose_json reader counts combining marks as part of a plain word,
as #544 now does.
@chhoumann

Copy link
Copy Markdown
Owner Author

PASS+NOTES
Verified head c69d365 by an independent verifier (GPT 6 Astra).

  • Compared master f457c5b0 and this head in isolated Obsidian 1.14.4 on Linux. Ran 24 baseline and 40 head source cases through Settings mouse/keyboard, file menus and palette commands. Used public/private fixture feeds on port 35018, a 70-minute MP3 and a 40-minute M4A, labels on/off, copied drivers and fake keys only.
  • All 302 captured uploads decode within their model caps. New models use five MP3 parts and eighteen M4A parts; all 100 offsets match independently measured durations. Timestamp options, omitted labels and model-specific notices pass. All 24 ordinary Grok/Scribe/MAI notes are byte-identical to master.
  • Replayed all 108 distinct archived answers. All 677 saved replay notes preserve word characters and add only punctuation after Markdown unescaping; non-plain answers retain baseline behavior. Earlier doubled-mark failures pass. Fresh cases cover composed/decomposed Latin, CJK, Hindi, Thai, Arabic, mixed scripts, numbers, quotes, compounds, filler, repetition and symbols. Supplementary generated/Unicode guard probes also pass.
  • Cached real Qwen answers retain all 2,262/3,785 words and restore 301/311 and 612/616 ASCII punctuation marks. Both 17,500-word notes retain every word; single fixture runs took 2.66/2.54 seconds including driver overhead. Only cached-answer timestamps were scaled.
  • Withdrawing either new model marks it unavailable, sends Grok with a warning and hides it after another selection. Voxtral saves its timed segments despite differing top-level text. Untimed Qwen/Voxtral answers write no note. Direct-command replays of earlier failures also pass.
  • Nonblocking notes: Well, we’ll see. can become we’ll, see.; separator-insensitive numeric matching can likewise place a comma on the wrong occurrence. Arabic ، and ؟ are not restored. The accepted 210-token filler limit remains. These cases preserve the supplied words.
  • All requested checks pass: npm run lint, npm run format:check, npm run typecheck, npm run build, npm run size, npm run test (2,946 tests), and npm run test:tooling (56 tests). Bundle size is exactly 330,546 raw / 108,180 gzip bytes. Commands used fnm exec --using 22.23.2; that installed entry reports v22.23.4.
  • Inspected screenshots, notices, requests, notes on disk and persisted settings. Every dev:errors report is clear. No tracked files changed. The fixture server and this worktree's Obsidian are stopped; the dry run reports pids: none. Evidence and copied drivers are in 6416f5d1-6392457b509bf5e4/.obsidian-e2e-artifacts/review544/, starting with verification.md.

@chhoumann
chhoumann merged commit 264ab15 into master Nov 16, 2026
18 checks passed
@chhoumann
chhoumann deleted the feat/transcription-openrouter-models branch November 16, 2026 16:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant