Groq Whisper Dictation in OpenTypeless: Setup and Error Checks
This walkthrough covers dictation through the built-in Groq Whisper route in OpenTypeless v1.1.61. It is for checking your own microphone-to-text workflow. It contains no measured speed or accuracy results. For the fixed endpoint, model and credential reference, use the linked provider documentation.
What the built-in route does
The released route uses whisper-large-v3-turbo and the Groq audio transcriptions endpoint. OpenTypeless collects audio during recording and uploads a WAV when recording ends; this route does not stream partial transcripts into your text field. Groq API features described in vendor documentation are not automatically settings available in this app.
Before you start
- Use a Groq account and API key that can access speech transcription. Check your current account limits and data terms in the official console.
- Choose a non-sensitive test sentence, a quiet microphone and a disposable text field.
- Keep AI polishing off for the first recording so an LLM cannot mask a recognition error.
- Check microphone access and the relevant insertion permissions for your operating system.
Configure the key, then make a real recording
- Open Settings → Speech Recognition and choose Groq Whisper.
- Save your Groq API key and click Test. Never put a key in a screenshot, issue or public recording.
- Treat a green connection result as an HTTP access check. In v1.1.61, Test uploads a small silent WAV; it does not listen to your microphone or check transcript accuracy.
- Select the recording language when appropriate, then start and stop a short recording using your configured shortcut.
- Review the raw transcript in a disposable destination: names, numbers, negation, language and punctuation. Confirm that text reached the intended field before using it in normal work.
Use a reference sentence that can expose errors
Illustrative reference: “The ticket is TM-482. We might release version 1.1.61 on Friday if the tests pass; do not promise the customer a release date.” This is a proposed test sentence, not a recording or benchmark result. Compare the transcript with what you actually said; mixed-language names and identifiers need separate samples.
Choose language and understand dictionary limits
The built-in route sends the selected language when it is not set to multilingual mode. If recognition uses an unexpected language, check that setting and retest the same spoken sentence. In this release, the Groq route has no configured extra prompt field: adding a word to the OpenTypeless dictionary does not send it as a Groq transcription prompt or retrain Whisper. The dictionary can inform the later LLM prompt when polishing is enabled.
Add polishing only after raw transcription works
Ordinary dictation can return raw STT text with polishing disabled. To try polishing, configure a usable LLM route, enable it and repeat the same recording. Compare raw and polished drafts: did the LLM preserve TM-482, 1.1.61, “might”, “if” and “do not”? A smoother sentence is not a successful correction if it changes those facts. Selected-text editing and generation have different LLM requirements; this raw-dictation check does not validate them.
Locate the failure before changing providers
- Authentication error: check the selected Groq route and saved credential. Use the provider error text; do not change the fixed model or endpoint to hide an invalid key.
- Rate-limit response: inspect current account limits and the displayed error before retrying. A website cannot promise free capacity or an unlimited quota.
- Test passes, recording fails: inspect microphone access, captured audio, network access and recording size. The connection test did not validate those stages.
- Wrong transcript or language: replay your sample, verify the input language and microphone, and compare raw STT before editing the polish prompt.
- Good transcript, wrong final wording: inspect the LLM route and prompt. Retry with polishing off to isolate the change.
- Text exists but is not inserted: check the focused destination and operating-system insertion permissions. Verify recovery or undo in a disposable field.
Measure the workflow you will actually use
Record the app version, language, STT route, LLM state, sample text and test date. Repeat comparable samples and record time from stopping the recording to an editable result, remaining errors and correction time. Keep those times separate from time spent speaking and from the Settings connection test. Compare the total work with typing the same task; use your own samples, speaker and network when making the decision.