Skip to content
EchoFlow

MODELS

Choose your brain.

EchoFlow uses two kinds of model: a speech model that understands what you say, and an optional AI model that tidies how it's written. Both run on your PC.

How EchoFlow uses them

  1. 01Your voiceYou hold Win + Alt and speak.
  2. 02Speech modelTurns what you said into a transcript.
  3. 03TranscriptYour words, with punctuation.
  4. 04 · OPTIONALAI refinementOptional. Tidies grammar, phrasing and tone.
  5. 05Text insertedTyped where your cursor is.

Speech recognition always comes first. AI refinement is an optional second step: with Direct Speech, EchoFlow inserts the transcript without AI rewriting, so it's quicker.

Once your models are downloaded, both steps run locally on your PC.

Speech models

EchoFlow uses OpenAI's Whisper speech models. Smaller models are faster and lighter; larger ones are more accurate but bigger downloads that need more of your PC's resources.

  • Tiny

    75 MB

    Quick notes in clear English, especially on older PCs.

    SpeedFastest
    AccuracyBasic
  • Base

    145 MB

    IN INSTANT VELOCITY

    Everyday clear English with fast responses.

    SpeedFast
    AccuracyGood
  • Small

    480 MB

    IN BALANCED PRECISION

    A balance of speed and accuracy for most PCs.

    SpeedBalanced
    AccuracyBetter
  • Medium

    1.5 GB

    Better accuracy, including for languages other than English.

    SpeedSlower
    AccuracyHigh
  • Large V3

    3.1 GB

    BEST FOR PUNJABI

    The highest accuracy, including strong accents, Hinglish and Punjabi. Best with a dedicated graphics card.

    SpeedSlowest
    AccuracyHighest

AI models

When AI refinement is on, an AI model running on your PC tidies the transcript in your chosen writing style. Smaller models respond faster and use less memory; larger ones can handle more nuance but are slower, especially without a graphics card.

Auto Brain

NOT A SEPARATE DOWNLOAD

Auto Brain isn't another model. It picks one of the AI models you've already downloaded for each dictation, based on your language, what you're writing and whether GPU acceleration is on. For English, it uses lighter models for short messages and larger ones for longer emails or technical writing. For other languages, it prefers the largest model you have. With GPU acceleration off, it favours Phi-3 Mini or Gemma 2. It needs at least one AI model downloaded.

  • Gemma 2 2B · Google

    1.6 GB

    IN INSTANT VELOCITY

    The lightest and quickest. Good for short messages and lower-memory PCs.

    Resource useLightest

    Licence: Gemma Terms of Use · you accept it before download

  • Phi-3 Mini · Microsoft

    2.4 GB

    IN BALANCED PRECISION

    A balanced choice with a modest memory footprint.

    Resource useLight

    Licence: MIT

  • LLaMA 3 8B · Meta

    4.9 GB

    Larger, for longer or more involved writing, such as detailed emails.

    Resource useHeavier

    Licence: Meta Llama 3 Community License · you accept it before download

  • Mistral NeMo 12B · Mistral AI

    7.1 GB

    The largest, best suited to PCs with a dedicated graphics card.

    Resource useHeaviest

    Licence: Apache 2.0

Setup profiles

During setup, EchoFlow checks your PC and recommends one of two combinations. You can choose either, and change models later in Settings › Speech Engine.

  • Instant Velocity

    about 1.9 GB

    Whisper Base + Gemma 2 2B

    Faster responses with a smaller memory footprint. EchoFlow recommends it for most PCs: any PC without an NVIDIA graphics card, or with less than about 8 GB of RAM.

  • Balanced Precision

    about 2.9 GB

    Whisper Small + Phi-3 Mini

    Higher accuracy and more detailed formatting. EchoFlow recommends it when it finds an NVIDIA graphics card and at least about 8 GB of RAM.

Instant Velocity includes Gemma 2, so EchoFlow asks you to accept Gemma's licence before downloading it.

Which should I choose?

Let EchoFlow decide
Keep the setup EchoFlow recommends. For the AI model, Auto Brain picks among the models you've downloaded.
The fastest, lightest setup
Instant Velocity, or a small speech model such as Tiny or Base with Gemma 2.
Better transcription accuracy
Move up one speech-model size: Small, then Medium or Large V3.
I dictate in Punjabi
Large V3. Smaller models are less reliable for Punjabi.
No AI rewriting
Turn on Direct Speech. Dictation only needs the speech model; the AI model isn't used.
A modest or older PC
Instant Velocity. A graphics card is optional: EchoFlow runs on the CPU.

Downloaded once, run on your PC

  • The EchoFlow installer is only about 11 MB because models download separately, when you choose them.
  • Model files are stored on your PC. Speech recognition and AI refinement run there too, and your dictation isn't sent to EchoFlow or a cloud AI service.
  • Once the models you need are installed, everyday dictation works offline. Downloading a model needs an internet connection.
  • Download, switch or delete models at any time in Settings › Speech Engine.

Full details are in our Privacy Policy.

Graphics card or CPU

  • A graphics card is optional. EchoFlow runs on the CPU.
  • With a supported graphics card, EchoFlow speeds up both models locally through Vulkan. This has been verified on NVIDIA graphics cards.
  • If the graphics card can't be used, EchoFlow switches to the CPU automatically. Larger models are slower on the CPU.

For memory and storage, see the full system requirements.

Model licences

  • EchoFlow itself is covered by the EchoFlow Terms. The models are made by other organisations and stay under their own licences.
  • EchoFlow shows each model's licence in Settings › Speech Engine. Gemma 2 and Meta Llama 3 ask you to accept their licences before EchoFlow downloads or uses them.
  • The Terms list every model and its licence.

More information

  • Download: system requirements and installation.
  • Help: choosing models, languages and troubleshooting.
  • Privacy Policy: how EchoFlow handles information.
  • Terms: the rules for using EchoFlow.