Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

docs: release v0.10.6-alpha - single-owner background work, sliced uploads, email notifications

Murtaza Nasirc670292efbb5
diff --git a/README.md b/README.md
index 6dd578a8..73aa5efa 100644
--- a/README.md
+++ b/README.md
@@ -9,7 +9,7 @@
<a href="https://www.gnu.org/licenses/agpl-3.0"><img alt="AGPL v3" src="https://img.shields.io/badge/License-AGPL_v3-blue.svg"></a>
<a href="https://github.com/murtaza-nasir/speakr/actions/workflows/docker-publish.yml"><img alt="Docker Build" src="https://github.com/murtaza-nasir/speakr/actions/workflows/docker-publish.yml/badge.svg"></a>
<a href="https://hub.docker.com/r/learnedmachine/speakr"><img alt="Docker Pulls" src="https://img.shields.io/docker/pulls/learnedmachine/speakr"></a>
- <a href="https://github.com/murtaza-nasir/speakr/releases/latest"><img alt="Latest Version" src="https://img.shields.io/badge/version-0.10.5--alpha-brightgreen.svg"></a>
+ <a href="https://github.com/murtaza-nasir/speakr/releases/latest"><img alt="Latest Version" src="https://img.shields.io/badge/version-0.10.6--alpha-brightgreen.svg"></a>
</p>
<p align="center">
@@ -212,11 +212,19 @@ Complete documentation is available at **[murtaza-nasir.github.io/speakr](https:
- [Troubleshooting](https://murtaza-nasir.github.io/speakr/troubleshooting) - Common issues and solutions
- [FAQ](https://murtaza-nasir.github.io/speakr/faq) - Frequently asked questions
-## Latest Release (v0.10.5-alpha)
+## Latest Release (v0.10.6-alpha)
+
+**Background work now runs once per installation instead of once per worker, plus large uploads through restrictive proxies and opt-in email notifications.** Speakr's startup code runs in every process that imports the application, so with the three gunicorn workers the image ships, each of them started its own copy of the background machinery. For the job queue that meant a process booting while another was mid-transcription could un-claim that live job and send the same audio to the ASR service again, once per worker, with the transcript still arriving correctly so nothing looked wrong (#384, reported and fixed by @jagd700). The same pattern affected the webhook dispatcher, the watch-folder monitor and the retention and cleanup schedulers. One process is now elected to own each of these, the rest serve requests and enqueue normally, and raising `--workers` for request capacity is safe.
+
+Files larger than a reverse proxy's request body limit can now be uploaded, sent in slices through the existing session endpoints and reassembled server-side before going through the ordinary ingestion path, with resume across a dropped connection (#381, contributed by @fcatuhe). Cloudflare's 100 MB body cap on most plans is well under a normal two-hour recording, and the only workaround was routing uploads around the proxy.
+
+Each user can now opt in to an email when a recording finishes transcribing or fails (#386), using the same SMTP configuration as verification and password resets. Speakr also detects when the transcription backend starts returning voice embeddings from a different model, which otherwise makes existing voice profiles silently stop matching with nothing in the logs to explain it (#380). Also fixed: a chat panel restored from a larger display could strand its close button off-screen (#383), the webhook completion payload never actually contained the duration field it documented, and user-supplied text in outgoing email is now escaped and the logo travels with the message rather than being linked. Database columns migrate automatically; no configuration changes are required. **Full release notes on the [GitHub release page](https://github.com/murtaza-nasir/speakr/releases/tag/v0.10.6-alpha).**
+
+### v0.10.5-alpha (previous release)
**A database migration fix for long-running installations.** A migration added with SSO support rebuilt the whole user table to change one column, which on databases upgraded from very old versions could fail partway, strand a temporary table and repeat the same error on every startup (#379). Worse, had that rebuild ever completed it would have restored the table from a column list frozen years earlier, discarding later settings such as token budgets, verification state and transcription hints. It now alters only the column it needs, leaving every other column, index and value untouched, and clears the stranded table once it confirms the real one is intact. Startup migrations are additionally serialised across worker processes, grouped so one failure cannot skip the rest, and reported clearly instead of as a single warning line. Upgrading is routine and no configuration changes are required. **Full release notes on the [GitHub release page](https://github.com/murtaza-nasir/speakr/releases/tag/v0.10.5-alpha).**
-### v0.10.4-alpha (previous release)
+### v0.10.4-alpha
**An agentic Inquire beta, reorganized tag and folder management, and upload workflow improvements.** Inquire gains an opt-in agent mode (`ENABLE_INQUIRE_AGENT=true`) that researches across your recordings with search, listing, and reading tools, shows a live activity timeline, and answers with numbered citations that click through to the exact moment in the recording; each user controls whether their summaries and notes are available to it. Search indexing now chunks along whole speaker turns with timestamps, which is what makes citation deep links land precisely. Tag and folder management merge into one account tab with a single shared card design, and both can now be created directly from the upload dialog with the full editor, including a proper localized language dropdown. The upload button now closes the dialog after queuing, with a new "Upload & Add More" button for the previous behaviour. A community-contributed OpenASR connector adds another local transcription option. Fixes cover the iOS locked-screen recording timer, merged-audio decoding on ASR pipelines, media playback through the service worker, SSO providers with minimal ID tokens, email deliverability headers, and the Identify Speakers save button. Database columns migrate automatically; no configuration changes are required. **Full release notes on the [GitHub release page](https://github.com/murtaza-nasir/speakr/releases/tag/v0.10.4-alpha).**
diff --git a/VERSION b/VERSION
index 3c8b8db5..7e9cfa34 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-v0.10.5-alpha
+v0.10.6-alpha
diff --git a/docs/admin-guide/system-settings.md b/docs/admin-guide/system-settings.md
index 2cbe0af2..9a044369 100644
--- a/docs/admin-guide/system-settings.md
+++ b/docs/admin-guide/system-settings.md
@@ -166,6 +166,12 @@ Speakr uses separate job queues for transcription and summarization to prevent s
Jobs are persisted to the database and survive application restarts. If Speakr restarts while jobs are processing, they automatically resume from where they left off.
+**These are threads, not processes.** `JOB_QUEUE_WORKERS` controls how many transcriptions run at once inside the one process that owns the queue, which is the right shape for this work: a worker spends its time waiting on an ASR endpoint or an LLM API rather than using CPU, so more threads means more concurrency without more memory.
+
+They are a separate thing from gunicorn's `--workers`, which controls how many processes serve HTTP. The container runs three of those by default. Because Speakr's startup code runs in every process that imports the app, background work is claimed by exactly one of them: the others serve requests and enqueue jobs normally, but run no workers, no watch-folder monitor and no schedulers. Before v0.10.6-alpha every process ran its own copy, which could send the same recording to the ASR service once per gunicorn worker. Raising `--workers` for more request capacity is safe.
+
+If the process that owns the background work is recycled or killed, the replacement claims it on startup and any job that was interrupted is recovered.
+
### Folders Feature
**ENABLE_FOLDERS**: Enable the folders organization feature. When `true`, users can create folders to organize recordings with per-folder custom prompts and ASR settings. Default: `false`.
diff --git a/docs/features.md b/docs/features.md
index c57e56bc..2b87ff19 100644
--- a/docs/features.md
+++ b/docs/features.md
@@ -127,6 +127,8 @@ Build and maintain a library of speaker profiles that persist across recordings.
**Voice Profile Recognition**: The system builds voice recognition profiles from speaker identifications across multiple recordings. These profiles use voice embeddings supplied by the transcription backend (256-dimensional with WhisperX, though nothing in Speakr requires that size) to suggest speaker matches when processing new recordings, making speaker identification faster and more consistent. The voice matching system calculates similarity scores to help identify speakers across recordings, even when audio quality varies. Voice profiles are automatically updated as more recordings are processed, improving recognition accuracy over time.
+**Email Notifications**: Each user can opt in to an email when one of their recordings finishes transcribing, or fails, from their [account preferences](user-guide/settings.md#notifications). It uses the same SMTP configuration as email verification and password resets, and is off by default. Set `APP_BASE_URL` so the message can link back to the recording. See [Email Setup](admin-guide/email-setup.md#processing-notifications).
+
**Embedding Model Changes**: Voice profiles are only comparable with embeddings from the model that produced them. If the transcription backend starts returning embeddings from a different model, whether because the endpoint was repointed or because the service upgraded itself, existing profiles stop matching anything. On its own this failure is silent: matching simply returns nothing.
Speakr detects it by sending a short fixed clip to the backend and comparing the embedding it gets back against the one recorded when voice matching first worked. The check runs when the transcription configuration changes, and on demand from the admin area. When it detects a change, Speakr reports it and keeps the original reference, so restoring the previous backend clears the warning by itself. If the change was intended, rebuild the affected voice profiles and then re-baseline to accept the new model. A backend that is simply unreachable is reported as a failed check rather than a changed model.
diff --git a/docs/user-guide/settings.md b/docs/user-guide/settings.md
index 4cd0cd51..5085d62a 100644
--- a/docs/user-guide/settings.md
+++ b/docs/user-guide/settings.md
@@ -133,6 +133,14 @@ This setting is not the same as **Naming Templates** (Account → Templates), wh
**Show timestamps in simple view.** When enabled, a compact `mm:ss` (or `h:mm:ss` for long recordings) timestamp appears next to each speaker label in the simple transcript view, making it easier to navigate long meetings. The bubble view is unaffected. Off by default to keep the existing aesthetic for users who prefer a clean read.
+### Notifications
+
+**Email me when a transcription finishes.** When enabled, Speakr sends a message to your account email address as each recording finishes transcribing, and again if one fails permanently. The message names the recording, its length and how long transcription took, and links straight to it. Summaries do not send mail, since a summary follows a transcript you have already been told about.
+
+Off by default. It is most useful for long recordings you would rather not sit and wait for; less so if you upload many files at once, because each one sends its own message.
+
+The setting only appears when your administrator has configured email delivery and your account has an address Speakr can reach. Accounts created through single sign-on whose provider did not supply an email address do not see it, because there is nowhere to deliver to.
+
### Transcript Editor
**Auto-save edits in the transcript editor.** When enabled, edits made in the transcript editor are saved automatically a couple of seconds after you stop typing, with a small `Saved` indicator to confirm. When disabled (the default), saves are explicit via the Save buttons or the `Ctrl+S` keyboard shortcut. Auto-save is useful for long editing sessions where the risk of losing work outweighs the predictability of explicit writes.