[PR #1392] [CLOSED] docs: add tested LocalAI speech-to-text recipe #1357

New issue

Closed

opened 2026-10-18 14:19:06 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-18 14:19:06 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/open-webui/docs/pull/1392
Author: @Nathan-Roll1
Created: 9/17/2026
Status: ❌ Closed

Base: main ← Head: codex/localai-orukeet-stt-guide

📝 Commits (1)

d80bb65 docs: add tested LocalAI speech-to-text recipe

📊 Changes

2 files changed (+116 additions, -0 deletions)

View changed files

➕ docs/features/chat-conversations/audio/speech-to-text/localai-stt-integration.md (+115 -0)
📝 docs/features/chat-conversations/audio/speech-to-text/stt-config.md (+1 -0)

📄 Description

Affiliation correction: I am the founder of Oruk, and Orukeet is our model. That relationship should have been explicit in the original description. The contribution-standard checkbox below is now unchecked. This PR remains closed.

Summary

Add a tested LocalAI speech-to-text recipe using its existing Orukeet gallery model. It covers the compatible API URL and key, multipart requests, current admin settings, Docker networking, persistent-setting overrides, and the user's Default speech setting. The change is limited to one guide and a link from the STT setup page; it adds no runtime code, dependencies, telemetry, or chat-provider changes.

LocalAI already installs this model from Hugging Face at a pinned GGUF revision with SHA-256 verification. Its necessary model download uses normal Hugging Face download accounting; cached recognition makes no extra counting requests. The guide identifies the weights' license and final-recording behavior.

No related issue. This documents the existing OpenAI-compatible STT path and LocalAI gallery entry.

Checklist

I have reviewed the relevant documentation and matched the existing style.
This PR meets Open WebUI's contribution standards: it is accurate, relevant to users, narrowly scoped, maintainable, and not promotional content, advertising, lead generation, SEO placement, or a request to list a product, service, provider, integration, gateway, tool, or company primarily for visibility.
I understand that PRs that do not meet these standards may be closed without review and will not be merged. Repeated, low-quality, off-topic, promotional, or intentionally misleading submissions may result in the contributor being blocked from future participation in Open WebUI repositories.

Notes for reviewers

Validated on macOS arm64 with Open WebUI 0.11.3, LocalAI 4.9.0 and the gallery's Q8 model:

The real LocalAI backend/model install commands succeeded, reusing checksum-verified weights. The authenticated Open WebUI transcription endpoint returned identical nonempty text on two runs each of English, German, Spanish and French WAVs; 0.1-, 1- and 5-second silence returned empty text. An unsupported upload returned 400, then a valid upload still succeeded.
A WebM/Opus upload passed Open WebUI's conversion and the LocalAI endpoint. Chromium's actual Voice Input and Confirm recording controls, fed the English WAV as a simulated microphone, put the transcript in the chat input. The real admin Audio page displayed and saved the settings. Physical microphone hardware and a container deployment were not tested.
npm run build passed for all 426 documentation pages; the new Markdown passes Prettier and git diff --check is clean.

For context, the published NeMo evaluation reports 9.85% versus 11.01% pooled FLEURS WER across 20,146 recordings in 25 languages. A separate 120-clip transcribe.cpp Q8 sample scored 5.34% versus 5.14%. Those are different runtime evaluations; neither establishes an accuracy or speed advantage through Open WebUI or this LocalAI backend. The checks above establish integration and repeatability. The guide includes the required community-support notice.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/open-webui/docs/pull/1392 **Author:** [@Nathan-Roll1](https://github.com/Nathan-Roll1) **Created:** 9/17/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `codex/localai-orukeet-stt-guide` --- ### 📝 Commits (1) - [`d80bb65`](https://github.com/open-webui/docs/commit/d80bb653728daf90021a3b53ce5be409b60734d1) docs: add tested LocalAI speech-to-text recipe ### 📊 Changes **2 files changed** (+116 additions, -0 deletions) <details> <summary>View changed files</summary> ➕ `docs/features/chat-conversations/audio/speech-to-text/localai-stt-integration.md` (+115 -0) 📝 `docs/features/chat-conversations/audio/speech-to-text/stt-config.md` (+1 -0) </details> ### 📄 Description **Affiliation correction:** I am the founder of Oruk, and Orukeet is our model. That relationship should have been explicit in the original description. The contribution-standard checkbox below is now unchecked. This PR remains closed. ## Summary Add a tested LocalAI speech-to-text recipe using its existing Orukeet gallery model. It covers the compatible API URL and key, multipart requests, current admin settings, Docker networking, persistent-setting overrides, and the user's Default speech setting. The change is limited to one guide and a link from the STT setup page; it adds no runtime code, dependencies, telemetry, or chat-provider changes. LocalAI already installs this model from Hugging Face at a pinned GGUF revision with SHA-256 verification. Its necessary model download uses normal Hugging Face download accounting; cached recognition makes no extra counting requests. The guide identifies the weights' license and final-recording behavior. ## Related issue or discussion No related issue. This documents the existing OpenAI-compatible STT path and LocalAI gallery entry. ## Checklist - [x] I have reviewed the relevant documentation and matched the existing style. - [ ] This PR meets Open WebUI's contribution standards: it is accurate, relevant to users, narrowly scoped, maintainable, and not promotional content, advertising, lead generation, SEO placement, or a request to list a product, service, provider, integration, gateway, tool, or company primarily for visibility. - [x] I understand that PRs that do not meet these standards may be closed without review and will not be merged. Repeated, low-quality, off-topic, promotional, or intentionally misleading submissions may result in the contributor being blocked from future participation in Open WebUI repositories. ## Notes for reviewers Validated on macOS arm64 with Open WebUI 0.11.3, LocalAI 4.9.0 and the gallery's Q8 model: - The real LocalAI backend/model install commands succeeded, reusing checksum-verified weights. The authenticated Open WebUI transcription endpoint returned identical nonempty text on two runs each of English, German, Spanish and French WAVs; 0.1-, 1- and 5-second silence returned empty text. An unsupported upload returned 400, then a valid upload still succeeded. - A WebM/Opus upload passed Open WebUI's conversion and the LocalAI endpoint. Chromium's actual Voice Input and Confirm recording controls, fed the English WAV as a simulated microphone, put the transcript in the chat input. The real admin Audio page displayed and saved the settings. Physical microphone hardware and a container deployment were not tested. - `npm run build` passed for all 426 documentation pages; the new Markdown passes Prettier and `git diff --check` is clean. For context, the [published NeMo evaluation](https://huggingface.co/oruk/orukeet#evaluation) reports 9.85% versus 11.01% pooled FLEURS WER across 20,146 recordings in 25 languages. A separate [120-clip transcribe.cpp Q8 sample](https://huggingface.co/oruk/orukeet/blob/main/transcribe-cpp/README.md) scored 5.34% versus 5.14%. Those are different runtime evaluations; neither establishes an accuracy or speed advantage through Open WebUI or this LocalAI backend. The checks above establish integration and repeatability. The guide includes the required community-support notice. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-18 14:19:06 +00:00