Set up speech recognition

Speech recognition turns your voice into text in the message box. Before the microphone can record, DataMind Oda prepares its recognition models on the computer and uses the service and language you choose. These controls live in the Voice input plugin's details, which you open from the Plugins page; how recording and dictation themselves work is covered in Voice input.

When Oda asks you to set up speech recognition

Voice input is part of a plugin you switch on first (Plugins and bundles).

After you switch it on, Oda checks the computer for speech models. If the local models are missing, a prompt appears:

The title reads Set up voice input before recording.
The body says Voice input is enabled. First use requires downloading and preparing local recognition models. Open the plugin details to review disk space, memory and time estimates before starting setup.
The buttons are Later and Go to setup.

Go to setup only opens the preparation details; nothing is downloaded until you press the download button there. Later closes the prompt and leaves voice input switched on, so you can prepare the models whenever it suits you. When intact models are already on the computer, the prompt does not appear and nothing is downloaded again.

Until recognition is ready, the microphone in the message box stays disabled and its tooltip reads Prepare recognition in the voice plugin settings first.

Prepare the models

Open the preparation details. Choose Go to setup in the prompt, or open the Voice input plugin from Plugins and find its card.
Expand the card. It starts collapsed; expanded, it shows Preparation steps.
Press Download and prepare.
Watch the steps run: Check local resources, Prepare recognition model, Prepare voice detection model, Verify model files, and Load with the model's name. Each step shows one of Not started, In progress, Completed, Failed or Cancelled. While a step works, its line counts the wait (Waiting 5 seconds) and the download shows a progress bar with its running totals.
Wait for the summary to report readiness: Local resources are prepared. Recording wakes the recognizer., Local speech recognition is ready., or, with a cloud service, Cloud speech recognition is ready..

Before you start, the card explains the download: Local models will be downloaded to the computer running DataMind Oda. No Python or compiler is required. It also shows three estimates for the local model: Disk space (about 5 GB for the models, the runtime and download caches), Memory (about 2 GB with the model loaded, and recognition can use more) and First setup (roughly 5 to 30 minutes, depending on your network and machine). The figures are estimates: a slow network can take longer, downloads are reused, and an idle model releases its memory by default.

Cancel preparation stops a running preparation. It is hidden while the recognizer is waking and disabled while the cancellation itself is in progress. A cancelled preparation keeps its completed downloads, and if a preparation fails, Retry preparation starts it again and reuses the files that were already downloaded and verified.

Choose the service and language

The card holds two settings:

Recognition service chooses the service that recognises your speech. A local service runs on the computer running DataMind Oda; a cloud service is reached over the internet.
Recognition language lists the languages to listen for: Detect automatically, Chinese, English, Cantonese, Japanese and Korean. The list follows the service, so changing the service can change which languages you can pick.

Your choices are saved as you set them and persist for later sessions. While the card reports Connecting to speech recognition…, the settings and preparation controls are disabled until the connection returns.

Where your audio goes depends on the service:

With the local service, the card says Audio is recognized on the computer running DataMind Oda. First setup downloads models from Hugging Face by default, so that machine must be able to access Hugging Face and its model download services. If a proxy is needed, configure it on that machine.
With a cloud service, the card says Audio will be sent to the selected cloud service.
Local here means the computer running DataMind Oda, which can be a different computer from the one whose browser you use.

If saving a choice fails, the card shows Speech recognition failed: <message>; try the choice again.

When a download fails

If a preparation stops with a failure, the card shows what happened, a short line of advice for the cause, and two values:

Download source: <source> names where the models were being fetched from.
Error code: <code> names the error the download hit, when there is one.

Cause	The card shows	What to check
The download service cannot be reached	`Cannot download <resource>: connection to the download service failed.`	That the computer running DataMind Oda can reach the download source and its model download services; configure a proxy on that machine if needed, then retry.
The download address does not resolve	`Cannot download <resource>: the download address could not be resolved.`	The DNS and proxy settings of the computer running DataMind Oda, then retry.
The download times out	`The download of <resource> timed out.`	The network or proxy connection of that computer, then retry later. Verified files are kept.
The certificate check fails	`Cannot download <resource>: certificate verification failed.`	The system clock, trusted certificates and proxy settings of that computer, then retry.
The download host answers with an error	`Cannot download <resource>: the download service returned HTTP <status>.`	That the download address is available and the proxy can reach the download service, or retry later.
A file is incomplete or fails its check	`The download of <resource> is incomplete or failed verification.`	Nothing to check; retry the download. Other completed and verified files are kept.
The file cannot be saved	`Cannot save <resource>: insufficient disk space or write permission.`	Free disk space and write permission for the model directory on that computer, then retry.
Anything else	`Could not prepare <resource>.`	The network, disk space and model directory permissions of that computer, then retry.

A preparation that fails for a reason the card cannot tie to a download step is reported as Preparation failed: <message>.

Start the preparation again from the same card. If it keeps failing, contact your administrator and include the message, the download source and the error code.

Questions, answered

Where do I find the speech recognition settings?

In the Voice input plugin's details, on the Plugins page. The card there holds the preparation controls and the Recognition service and Recognition language lists, and the first-use prompt's Go to setup button takes you straight to it.

Why does the microphone say `Prepare recognition in the voice plugin settings first`?

The speech models are not prepared yet, so the microphone stays disabled and that is its tooltip. Prepare the models from the plugin's details; recording works once the card reports the recognizer is ready.

How do I know when speech recognition is ready?

The card's summary reports it: Local resources are prepared. Recording wakes the recognizer. for the local service, Local speech recognition is ready. while the recognizer is awake, or Cloud speech recognition is ready. with a cloud service.

What does `Connecting to speech recognition…` mean?

The app is not connected to the speech service, usually for a moment while it starts or reconnects. The settings and preparation controls are disabled until the connection returns, and the card updates by itself.

Can I change the recognition service or language later?

Yes. Your choices are saved as you set them and persist for later sessions. The language list follows the selected service, so a different service can offer a different set of languages.

Can I choose a cloud service instead of the local one?

If your app offers one, choose it under Recognition service. With a cloud service the card says Audio will be sent to the selected cloud service., and the local model's preparation no longer applies.

What should I check when a model download fails?

Read the advice line first: the card distinguishes a connection problem, a name-resolution problem, a timeout, a certificate problem, an HTTP error from the download host, an incomplete or failed file check, and not enough disk space. The table above lists what to check for each cause. Completed and verified files are kept, so you can start the preparation again from the card. If it keeps failing, contact your administrator and include the message, the download source and the error code.

The card shows `Preparation failed: <message>` or `Speech recognition failed: <message>`. What is the difference?

Preparation failed: <message> is the preparation itself stopping with an error the card could not tie to a download step. Speech recognition failed: <message> is an action not going through, such as starting or cancelling a preparation, or saving a service or language choice. In both cases, try the action again; the message names the reason.

I cancelled preparation. Will completed downloads be made again?

No. A cancelled preparation keeps what already downloaded (Preparation cancelled. Completed downloads are retained.), and Retry preparation reuses the files that were downloaded and verified.

What do the `Disk space`, `Memory` and `First setup` estimates cover?

Disk space covers the models, the runtime and the download caches; Memory covers the model once it is loaded, and recognition can use more; First setup is a time estimate that depends on your network and machine. For the local model the card asks for about 5 GB of disk space and about 2 GB of memory. The figures are estimates: a slow network can take longer.