Skip to main content

Transcription Settings

Which model, which language, and how to make the app get your vocabulary right. Start with the recommended model and change nothing else until something bothers you.

Model Selection​

Location: Settings > Transcription

Choose which AI model to use for speech recognition. The model picker is searchable: start typing a model name to jump straight to it, use the arrow keys to move through the list, Enter to pick, Escape to close.

Available Models​

Knowii Voice AI supports multiple transcription models with different characteristics. Choose based on your language needs, hardware capabilities, and accuracy requirements:

Parakeet Models​

:::tip Numbers coming out as words? If digit sequences appear as words ("one one two two" instead of "1122"), enable Write digit sequences as numbers in Settings > Transcription, or use Parakeet V2 for English dictation. :::

  • Parakeet V3 (GPU): the fastest choice on machines with a graphics card

    • Same accuracy and languages as Parakeet V3
    • Uses your graphics card when one is available — up to 3× faster transcription, and the model loads in a fraction of the time
    • Falls back to your processor automatically when no graphics card is available
    • 740 MB download
    • May write numbers as words, like Parakeet V3 below

  • Parakeet V3 (Recommended): the default choice for most people

    • Fast and accurate
    • Supports 25 European languages
    • Automatic language detection
    • 785 MB download
    • Excellent balance of speed and accuracy
    • Latest version with improved performance
    • May write numbers as words ("one one two two" instead of "1122"). If you often dictate numbers in English, use Parakeet V2 instead

  • Parakeet V2: the English-only alternative when you dictate numbers

    • Fast with good accuracy
    • English only
    • 697 MB download
    • Writes numbers, dates, and amounts as digits more consistently than V3 ("1122", "March 25, 2026")
    • The best choice for dictating account numbers, phone numbers, prices, or dates in English

Whisper Models (Multi-Language)​

Support 99 languages including English, Spanish, French, German, Chinese, Japanese, and many more. Since the next release the Whisper models use your graphics card when you have one (and your processor otherwise), and they are about half the download they used to be:

  • Tiny: Very fast, basic accuracy (46 MB, ~0.5GB RAM)
  • Small: Fast and quite accurate (270 MB, ~2GB RAM)
  • Medium: Accurate but slower (832 MB, ~5GB RAM)
  • Large V3 Turbo: Accurate but slow (886 MB, ~6GB RAM)
  • Large V3: Highest accuracy but slowest (1670 MB, ~10GB RAM)

Whisper Models (English-Only)​

Optimized specifically for English transcription:

  • Tiny (English only): Very fast (46 MB, ~0.5GB RAM)
  • Base (English only): Fast (85 MB, ~1GB RAM)
  • Small (English only): Fast and accurate (270 MB, ~2GB RAM)
  • Medium (English only): Accurate (832 MB, ~5GB RAM)

:::note Whisper models are downloaded again after updating The next release switches the Whisper models to a new file format (the same one Parakeet V3 (GPU) uses). A Whisper model you downloaded before shows as not downloaded after the update; download it again from this page — the old file is removed automatically so it does not take up space. Your model choice, language and prompt settings are kept. :::

Omnilingual Models​

Speak any of 1,600+ languages and the model figures out what you're saying automatically. Learn more about Omnilingual: Blog post | Research paper | GitHub | Demo video

  • Omnilingual 300M: Smaller and faster, ~350 MB download
  • Omnilingual 1B: Larger and more accurate, ~985 MB download

Why choose Omnilingual:

  • Supports over 1,600 languages, more than any other model
  • Automatically detects what language you're speaking
  • Great for multilingual users who switch between languages
  • Covers less common languages that no other model supports
  • No need to manually select your language

Note: Still experimental, so accuracy may vary depending on language and audio quality.

Moonshine Models​

Lightweight models designed for computers with limited resources (low memory or CPU). Learn more about Moonshine: Research paper | Introduction blog post | GitHub | Hugging Face

  • Moonshine Base (English): Better accuracy, ~77 MB
  • Moonshine Tiny (English): Fastest, smallest footprint, ~35 MB
  • Moonshine Base / Tiny (Arabic): Arabic transcription, ~77 MB / ~35 MB
  • Moonshine Base / Tiny (Chinese): Chinese transcription, ~77 MB / ~35 MB
  • Moonshine Base / Tiny (Japanese): Japanese transcription, ~77 MB / ~35 MB
  • Moonshine Base / Tiny (Korean): Korean transcription, ~77 MB / ~35 MB
  • Moonshine Base / Tiny (Ukrainian): Ukrainian transcription, ~77 MB / ~35 MB
  • Moonshine Base / Tiny (Vietnamese): Vietnamese transcription, ~77 MB / ~35 MB

Moonshine models run on the same engine as Whisper and Parakeet V3 (GPU): your graphics card when there is one, your processor otherwise. A Moonshine model downloaded before the switch needs to be downloaded again — the old files are removed for you at startup, and the new download is a single file a third of the size.

Moonshine characteristics:

  • 5-15x faster than Whisper Tiny/Base/Small/Medium models
  • Ideal for low-end machines with limited RAM or CPU
  • Much smaller memory footprint than other models
  • Less accurate than Parakeet or Whisper (trade-off for speed and size)
  • Each model variant supports only one language
  • Good option when other models won't run on your hardware

Supported languages:

  • English (Base and Tiny variants)
  • Arabic (Tiny variant)
  • Chinese (Tiny variant)
  • Japanese (Tiny variant)
  • Korean (Tiny variant)
  • Ukrainian (Tiny variant)
  • Vietnamese (Tiny variant)

How to Choose a Model​

For most users:

  • Start with Parakeet V3 if you speak European languages (recommended)
  • Use Parakeet V2 if you dictate in English and want numbers written as digits (account numbers, prices, dates)
  • Use Small Whisper model for other languages
  • Upgrade to larger models if accuracy is insufficient

For English-only users:

  • Use Small (English only) for best balance
  • Upgrade to Medium (English only) for higher accuracy

For multilingual use:

  • Parakeet V3 for European languages (automatic detection)
  • Small or Medium Whisper for global language support

Hardware considerations:

  • Machine with a graphics card: Parakeet V3 (GPU) is the fastest option
  • Very limited RAM (under 2GB free): Use Moonshine Tiny models
  • Limited RAM (2-4GB free): Use Moonshine Base or Whisper Tiny/Small models
  • Moderate RAM (4-8GB): Use Parakeet or Whisper Small/Medium models
  • Plenty of RAM (8GB+): Large models give the best accuracy
  • SSDs recommended for faster model loading

Downloading Models​

Models are downloaded from the Settings > Transcription page:

  1. Scroll to the Available Models section
  2. Click Download on your chosen model
  3. Wait for download and installation to complete
  4. Model automatically becomes available for selection

Note: You can download multiple models and switch between them at any time.

Using Your Own Model Files​

If you have a model file the app does not offer — a fine-tuned Whisper, a quantization you prefer, a model you converted yourself — you can use it without waiting for it to appear in the list:

  1. Turn on Use My Own Model Files under the model list
  2. Click Open Custom Models Folder and drop the file in
  3. Click Rescan (the list also refreshes on its own the next time it is opened)

The file shows up in Available Models with a Your file badge and works like any other model: activate it, dictate, deactivate. One format is recognised: .gguf files, which the engine identifies by content (Whisper, Parakeet, Moonshine, …). Whisper GGUF files for every size are published at huggingface.co/handy-computer.

Anything else in the folder is ignored — including the older ggml-*.bin Whisper files, which the next release no longer reads — and so are sub-folders. A few things to know:

  • The app never downloads, verifies or deletes these files. There is no Delete button for them — remove the file from the folder yourself and rescan.
  • The Language setting shows every language for a custom model, because the app cannot know which ones the file was trained on.
  • The switch is off by default so a stray file cannot quietly become a pickable model.

Language​

Location: Settings > Transcription

Select the language for speech recognition:

Auto Detect (Default)​

  • Automatically detects the spoken language
  • Works with multi-language models (Parakeet, Whisper multi-language)
  • Recommended for most users
  • No manual language selection needed

Manual Language Selection​

When using multi-language Whisper models, you can specify the language:

  • Improves accuracy if you always speak the same language
  • Reduces processing time slightly
  • Available languages depend on the selected model
  • Searchable dropdown with 99+ languages: just start typing the language name (for example, type "eng" to jump to English)

Reset Button: Click the reset icon to return to Auto Detect.

Model-Specific Behavior​

English-only models (.en variants):

  • Language automatically set to English
  • Cannot change to other languages
  • Language selector is disabled

Parakeet models:

  • Automatically detect language
  • Language selector is disabled
  • Supports 25 European languages

Omnilingual models:

  • Automatically detect language from 1,600+ languages
  • Language selector is disabled
  • No manual language selection needed

Multi-language Whisper models:

  • Auto Detect by default
  • Can manually specify language for better accuracy
  • Full language selector available

Supported Languages​

The available languages depend on your selected model:

Omnilingual Models (1,600+ languages):

  • Supports virtually every language in the world
  • Includes rare and less common languages not available in other models
  • Automatic language detection, with no manual selection needed

Parakeet V3 (25 European languages):

  • Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish

Whisper Models (99 languages):

  • All major world languages including English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Arabic, Chinese, Japanese, Korean, Hindi, and many more
  • See the full list in the language selector dropdown

Custom Words​

Location: Settings > Transcription

Add specialized vocabulary to improve transcription accuracy:

How It Works​

  • Add words that are frequently misheard or incorrectly transcribed
  • AI model prioritizes these words during transcription
  • Useful for technical terms, names, acronyms, and domain-specific vocabulary

Adding Custom Words​

  1. Type the word in the input field
  2. Press Enter or click Add
  3. Word appears as a tag below the input
  4. Word is immediately active for future transcriptions

Requirements:

  • Single words only (no spaces allowed)
  • Maximum 50 characters per word
  • No special HTML characters (<>"'&)
  • Case-sensitive (add variations if needed)

Removing Custom Words​

Click the X button on any word tag to remove it from your custom vocabulary.

If you add a word that's already in the list, the app tells you instead of adding it twice.

:::tip Custom Words vs. Word Replacements Custom Words are a hint to the AI model. They help it lean toward the right spelling, but they don't guarantee it. If a term is always transcribed the same wrong way, use Word Replacements instead for an exact, guaranteed fix. :::

Best Practices​

Good examples of custom words:

  • Technical terms: "Kubernetes", "PostgreSQL", "TypeScript"
  • Company names: "Anthropic", "OpenAI"
  • Product names: "iPhone", "GitHub"
  • Acronyms: "API", "SDK", "CI/CD"
  • Uncommon names: "Knowii", "Docusaurus"
  • Domain jargon: "transcription", "embeddings"

Tips:

  • Add variations if needed: "GitHub" and "Github"
  • Include both singular and plural if frequently used
  • Add words after noticing repeated transcription errors
  • Don't add too many words (20-30 is typically sufficient)
  • Remove words you no longer use frequently

Word Replacements​

Location: Settings > Transcription

Sometimes a word, name, or acronym is always transcribed the same wrong way. Word Replacements let you fix it with an exact, guaranteed find-and-replace that runs on every transcription.

How It Works​

  • Enter the text that keeps appearing (Replace this) and what it should become (With this)
  • The correction is applied automatically after each transcription
  • Matches whole words only and ignores capitalization, so "kubernetes", "Kubernetes", and "KUBERNETES" are all caught
  • Runs after Custom Words, so it has the final say on those stubborn terms

Adding a Replacement​

  1. Type the misheard text in the Replace this field
  2. Type the correct text in the With this field
  3. Press Enter or click Add
  4. The rule appears in the list below and is active immediately

You can leave the With this field empty to simply delete an unwanted word from your transcriptions.

Word Replacements vs. Custom Words​

  • Custom Words are a gentle nudge. They help the AI model prefer a spelling, but don't guarantee it
  • Word Replacements are an exact rule. If you know a term always comes out wrong, this fixes it every time

Good examples:

  • A brand name the model always misspells the same way
  • An acronym consistently written as separate letters
  • A common autocorrect-style slip you want reversed

Phonetic Replacements​

Location: Settings > Transcription

If you dictate in one language but use names or acronyms from another, the AI model sometimes writes them out phonetically in the wrong script. For example, dictating in Russian, the tool "N8N" might come out as the Cyrillic "эн восемь эн". Phonetic Replacements fix exactly these cross-language cases.

How It Works​

  • Enter the sound as it gets transcribed (Transcribed sound) and the correct spelling (Correct spelling)
  • The correction is applied early, before Custom Words, so it reaches names and acronyms that normal correction can't
  • Ideal for terms that appear in a different alphabet than the rest of your text

When to Use It​

  • Technical terms, product names, or acronyms written in the wrong script
  • Terms that mix letters and numbers (like "N8N") and get spelled out phonetically
  • Anything Custom Words and Word Replacements can't catch because it comes out in another alphabet

Tip: For same-language corrections, use Word Replacements instead. Reach for Phonetic Replacements when the misheard text appears in a different script.

Your dictionary works outside the app too

Custom Words, Word Replacements, Phonetic Replacements and the Whisper Prompt together are your dictionary. Other tools can now read it, starting with Knowii Video AI, and the transcribe command-line tool can use it when it transcribes a file. They only read it: nothing outside the app changes your settings. See Your app dictionary.

Whisper Prompt​

Location: Settings > Transcription

When using Whisper models, you can give the AI a short hint about how you'd like your text formatted. This is a great way to nudge Whisper toward proper punctuation, capitalization, or specific vocabulary.

How It Works​

  • Type a short instruction or sample text in the Whisper Prompt box
  • For example: "Add proper punctuation and capitalization."
  • The hint gently steers how Whisper writes your transcription
  • Your Custom Words are automatically combined with this prompt

Note: This setting only affects Whisper models. Other models (Parakeet, Omnilingual, Moonshine) ignore it. It is optional and empty by default.

Remove Filler Words​

Location: Settings > Transcription

Turn on Remove filler words to automatically clean up hesitations and stutters from your transcriptions, so your dictated text reads more like polished writing.

What It Cleans Up​

  • Removes hesitation sounds like "uh", "um", and "hmm"
  • Shortens accidental stutters, so "doc doc doc" becomes "doc"
  • Which filler words are removed depends on your transcription language

Custom Filler Words​

When the feature is on, you can optionally add your own filler words to remove:

  • Type a word and press Enter or click Add
  • As soon as you add any custom words, they replace the built-in list for your language
  • Leave the list empty to keep the built-in defaults

Off by default: This option is turned off so you always get verbatim transcriptions unless you choose otherwise. Turn it on if you prefer cleaner, more readable text.

Write Digit Sequences as Numbers​

Location: Settings > Transcription

Some AI models write spoken digits as words: saying "1122" can come out as "one one two two" (Parakeet V3 does this regularly in English). Enable Write digit sequences as numbers to automatically convert those runs into digits after transcription:

  • "one one two two" becomes "1122"
  • "One, one, seven, six, eight, two, one, one" becomes "11768211"

Safe by design:

  • Only runs of 3 or more digits in a row are converted, so normal sentences like "I have one thing" or "give me two minutes" are never changed
  • Numbers are never merged across sentence boundaries
  • Off by default. Enable it if you often dictate account numbers, phone numbers, or codes

Tip: For English dictation, Parakeet V2 also writes numbers as digits natively. See Model Selection.

Common Scenarios​

English-Only User​

Recommended settings:

  • Model: Small (English only) or Medium (English only)
  • Language: English (automatically set)
  • Custom Words: Add technical terms from your field

Multilingual European User​

Recommended settings:

  • Model: Parakeet V3 (or V2)
  • Language: Auto Detect (automatically set)
  • Custom Words: Add names and technical terms in your languages

Global Multilingual User​

Recommended settings:

  • Model: Small or Medium Whisper (multi-language)
  • Language: Auto Detect or specify your primary language
  • Custom Words: Add names and terms in your languages

Technical Professional​

Recommended settings:

  • Model: Medium or larger (higher accuracy needed)
  • Language: Auto Detect or your primary language
  • Custom Words: Extensive list of technical terms, APIs, frameworks

Content Creator​

Recommended settings:

  • Model: Large V3 or Large V3 Turbo (best accuracy)
  • Language: Specify language for consistency
  • Custom Words: Brand names, product names, catchphrases

Low-End Hardware / Limited Resources​

Recommended settings:

  • Model: Moonshine Tiny (your language) for minimal resource usage
  • Model: Moonshine Base (your language) for slightly better accuracy
  • Language: Selected automatically based on model variant
  • Custom Words: Keep to a minimum

Why Moonshine for limited hardware:

  • Runs on computers where Parakeet or Whisper models struggle
  • Much lower memory requirements than other models
  • Acceptable accuracy for basic transcription needs
  • Trade-off: Slower processing and lower accuracy than larger models

Available Models Section​

Location: Settings > Transcription > Available Models

This section shows all AI models that can be downloaded:

  • Model name and description: What the model is designed for
  • Download button: Install models you don't have yet
  • Model details: Size, languages supported, accuracy level
  • Status indicators: Downloaded, active, or available for download

Managing Models​

Downloading new models:

  1. Browse the Available Models list
  2. Click Download on the model you want
  3. Wait for download to complete
  4. Model appears in the Model selector dropdown

Switching models:

  1. Select different model from the Model dropdown
  2. Model loads automatically (first use has a delay)
  3. Model stays loaded until timeout (see Advanced Settings)

Removing models:

  • Models can be removed through the application data folder
  • See Application Data for details

Troubleshooting​

Transcription Accuracy Issues​

  1. Try a larger model: more accurate but slower
  2. Add custom words: for frequently misheard terms
  3. Specify language: instead of Auto Detect
  4. Improve audio quality: better microphone, less background noise
  5. Speak clearly: moderate pace, clear pronunciation

Wrong Language Detected​

  1. Manually select language: instead of Auto Detect
  2. Use language-specific model: English-only for English
  3. Speak more before stopping: the model needs enough audio to detect
  4. Check model supports your language: see the supported languages list

Model Won't Download​

  1. Check internet connection
  2. Verify sufficient disk space (see model size)
  3. Check firewall isn't blocking download
  4. Try downloading again
  5. Check Application Data folder permissions

Model Loading Slow​

  1. First load is slower: the model needs to load into memory
  2. Adjust unload timeout: see Advanced Settings
  3. Use a smaller model: faster loading time
  4. Upgrade to SSD: much faster model loading

Custom Words Not Working​

  1. Verify word was added successfully (appears as tag)
  2. Check spelling matches exactly how you pronounce it
  3. Add phonetic variations if needed
  4. Try the word in isolation to test recognition
  5. Use larger model for better custom word recognition
  6. If a term is always wrong in the same way, use Word Replacements for a guaranteed exact fix
  7. If the term comes out in a different alphabet (a different script), use Phonetic Replacements

Tips​

  • Start with recommended models and adjust based on your needs
  • Download multiple models for different use cases
  • Use Auto Detect unless you have accuracy issues
  • Add custom words as you encounter transcription errors
  • Larger models require more RAM but provide better accuracy
  • English-only models are faster than multi-language equivalents for English
  • Test different models to find the best fit for your voice and use case
  • Keep your most-used model loaded by adjusting unload timeout

Need Help?​

If you have questions about transcription settings, visit the Support page.