Repository navigation
Releases: jaysqvl/Jotist
Release list
v2.3.2
v2.3.1
v2.3.0
v2.2.0
v2.1.0
Saved profile selection
- Search saved profiles when adding a run, running again, or transcribing in bulk. Filter by language and sort by published WER, memory estimates, name, or recent updates.
- Remember the selected profile, language filter, and sort order in your browser for each account. Existing profiles and their saved settings are preserved.
v2.0.0
2.0.0 (2026-10-02)
- Install and run the CLI as
jotist. Configuration uses~/.jotist.yamland
JOTIST_*variables, and the watcher service isjotist-watcher. - New containers and archives use
jotist-serverand default tojotist.db.
Existing installations must retain their data and signing secret, migrate
configuration and scripts, and sign in again. There are no Scriberr aliases
or automatic fallbacks; follow the migration guide. - Jotist names now cover installers, CLI downloads, browser storage, webhooks,
runtime messages, and source modules.
[changes]
Advance the release manifest to 2.0.0 and explain the canonical Jotist installation and migration.
[notes]
This release deliberately removes Scriberr command aliases and legacy configuration discovery. Preserve existing data and signing secrets when migrating; publication validates the exact tagged source.
v1.9.0
1.9.0 (2026-10-02)
- Configure automatic recovery globally in Settings, with optional overrides per
transcription profile. - See jobs across recordings in one queue, including recording names, waiting
targets, and numbered stage progress. Expand the two queued rows to see all jobs. - Improve mobile recording controls, run selection, and model browsing. Detailed
run statistics and diagnostics are available in Settings → Statistics. - Use the complete transcription and diarization model selectors, with referenced
publisher benchmarks and local timing and GPU memory observations.
The queue now defaults to one simultaneous recording. Change the limit in
Settings; an existing QUEUE_WORKERS environment override still takes precedence.
Saved recordings, profiles, preferences, and run history are retained. Stage
numbers are recorded for new runs; older runs without those fields keep their
stage labels.
v1.8.0
Changes
- Search and filter saved runs with clearer profile, model, CPU/GPU, recovery and resource details.
- Let active transcription continue without a default duration cap, using shared inactivity monitoring and preserving temporary audio until completion.
- Focus the default picker on English meeting models and compare published multilingual results for exact checkpoints.
- Improve recognition and timing recovery, speaker batch handling and queue writes. Retire Voxtral Mini 4B Realtime from the supported catalog while keeping historical runs readable and Mini 3B available.
Qualification
Source checks include shared worker/output contracts and mandatory real staged Whisper CPU inference. Focused installed-runtime checks passed for Canary recognition/alignment on CPU and CUDA, DiariZen speakers on CPU and CUDA, BitNet CPU recognition/alignment and Granite PLUS CPU recognition.
Granite PLUS recognition completed a 3,588.816-second recording in 17,923.66 seconds (RTF 4.99). That long replay used an earlier harness with matching native parser/retry logic; a separate current-source public fixture passes. It is slow recognition evidence without external diarization or a human accuracy reference.
The MOSS hour-long CUDA replay still fails its bounded terminal timing recovery and remains unqualified. See the qualification record for exact evidence and limits.
New ordinary runs have no default execution duration cap. The conservative shared inactivity monitor defers termination when activity or telemetry is uncertain. Explicit caller and historical saved deadlines remain enforced.
v1.7.6
1.7.6 (2026-09-27)
[changes]
- Prevent shared Qwen word alignment from failing when a predicted timestamp falls beyond its audio window. Timestamp decoding selects only times within the supplied audio while preserving recognized text.
- Include the affected window number in alignment errors. This covers LocalASR, legacy Voxtral, and VibeVoice BitNet paths using the shared aligner.
[notes]
The shared implementation passed all 18 fix PR checks and a complete replay of the affected alignment workload on the NAS, preserving every recognized word.
Start a new run after upgrading. Existing failed runs and saved recognition text remain available; checkpoint reuse requires compatible runtime identities.