Voice to Text for Machine Learning Engineers

Your brain is tracking hyperparameters, loss curves, and feature transformations. Blurt lets you speak your experiment notes, model documentation, and Slack updates while your mind stays on the problem. Hold a button, say what you need, release. Text appears in Jupyter, VS Code, or Weights and Biases. No context switching. Just talk and type.

Free to start Works in Jupyter, VS Code, W&B No configuration needed
Download Blurt Free

The Typing Problem

Documenting why this architecture, not that one

You just spent two hours comparing transformer variants. The LSTM worked better for your sequence length, but you know you'll forget why in a week. Writing up the decision takes another 20 minutes of typing. Your hands are tired from coding. The documentation never gets written. Three months later, a teammate asks why you didn't use attention, and you have no answer.

Experiment notes that disappear into the void

You're on experiment 47. Learning rate 0.001 with batch size 32 gave interesting results, but you didn't write down why you tried it. Now you're staring at a Weights and Biases dashboard with 50 runs and no context for any of them. The run names say 'test_v3_final_really_final'. Your past self left no breadcrumbs.

Model cards nobody wants to write

The model is ready for deployment but needs a model card. Intended use, limitations, bias considerations, training data description. You could explain it all in five minutes out loud to a colleague. Typing it out? That's an hour of documentation you'll keep pushing to tomorrow. The model ships without proper documentation. Again.

Slack messages during training runs

Your model is training and a PM asks for a status update. You could type a quick reply, but your brain is monitoring loss curves and thinking about whether to add dropout. The context switch to compose a message pulls you out completely. You type 'going well', but that's not actually helpful to anyone. The detailed update never gets written.

Jupyter markdown cells stay empty

You know notebooks should be self-documenting. Every tutorial has beautiful markdown cells explaining the analysis. Your production notebooks have code blocks and nothing else. Adding explanations means more typing after already typing code. The analysis is correct but incomprehensible to anyone else. Including future you next quarter.

How It Works

Blurt works in every app ML engineers use — Jupyter, VS Code, Colab, Weights and Biases, Slack. Anywhere you can put a cursor.

1

Hold your hotkey

Press your chosen shortcut. A small indicator shows Blurt is listening.

2

Talk naturally

Say your experiment notes, architecture decision, or Slack reply. Blurt handles punctuation.

3

Release and done

Text appears at your cursor. No copying, no pasting, no extra steps.

Real Scenarios

Documenting architecture decisions in real-time

You just decided to use a residual connection instead of a highway network. Hold the button and speak: 'Chose ResNet-style skip connections over highway networks because our gradients were vanishing at layer 8. Highway networks add learnable gating which introduces more parameters we can't afford with our dataset size.' Decision documented in 10 seconds while the reasoning is still fresh.

Writing model cards that actually get written

The model needs documentation before handoff. Hold and talk through each section: 'This model predicts customer churn from behavioral data. Trained on 18 months of interaction logs. Performs best on users with at least 30 days of history. Known limitation: underperforms on users acquired through the mobile campaign due to different event schema.' Model card done in 3 minutes instead of an hour.

Explaining feature engineering choices

You engineered 47 features and need to document them. Instead of typing descriptions for each, hold the button: 'Days since last purchase log transformed to handle right skew. Capped at 365 because users inactive longer than a year show different behavior patterns. Combined with purchase frequency to create recency frequency score.' Documentation happens at the speed of thought.

Quick Slack updates during training

The PM asks how the new model is looking. Your training loss just plateaued and you're thinking about next steps. Hold, speak: 'Training loss plateaued at 0.3 after epoch 20. Going to try learning rate decay and adding augmentation. Will have results by end of day.' Detailed update sent in 5 seconds. Back to watching those curves.

Adding context to Jupyter notebooks

Your notebook has 30 code cells and zero markdown. Before sharing with the team, add explanations by speaking: 'This cell handles missing values by forward filling for time series columns and median imputation for static features. We use forward fill because temporal patterns matter for our prediction window.' Notebooks become shareable without an hour of typing.

Documenting failed experiments for future reference

The batch normalization experiment failed spectacularly. Before moving on, capture what happened: 'Batch norm after attention layers caused training instability. Loss oscillated between 2.3 and 4.1. Suspect issue with varying sequence lengths in each batch affecting batch statistics. Trying layer norm next.' Failed experiments become learning, not wasted time.

Why ML engineers choose Blurt over built-in dictation

Blurt macOS Dictation
Activation Single hotkey, instant start Click microphone icon or 'Hey Siri'
Speed Text appears in under 500ms 2-3 second delay before transcription
Technical vocabulary Handles ML terms like 'hyperparameter' and 'backprop' Often mangles technical terminology
Reliability Consistent accuracy across long sessions Often fails silently or requires restart

Frequently Asked Questions

Does Blurt work inside Jupyter notebooks?
Yes. Blurt works anywhere you can type on macOS. Click into a markdown cell or code comment in Jupyter, hold your hotkey, speak, and release. Works in JupyterLab, classic Jupyter, and VS Code notebooks.
Can Blurt handle ML-specific terminology?
Blurt handles most ML vocabulary well. Terms like 'hyperparameter', 'backpropagation', 'softmax', and 'convolutional' transcribe correctly. For very specialized terms or custom layer names, occasional edits might be needed.
Does it work with Weights and Biases and experiment tracking tools?
Yes. Blurt inserts text wherever your cursor is. If you're adding notes in W&B, MLflow, or any web-based tracking tool, Blurt works. Just place your cursor, hold, speak, release.
Can I dictate while a training run is happening?
Absolutely. Blurt runs independently as a menu bar app. Your training continues uninterrupted while you dictate notes, Slack messages, or documentation. No resource conflicts.
What about code comments in Python?
Blurt is great for code comments and docstrings. Place your cursor where the comment goes, hold your hotkey, explain what the function does, release. The text appears. You'll need to add the comment syntax yourself.
Does Blurt rewrite or clean up my words?
No. Blurt transcribes exactly what you say with proper punctuation and capitalization. It doesn't rewrite, summarize, or use AI to change your words. What you say is what appears.
Does Blurt work on Windows or Linux?
Blurt is macOS only. We focused on creating the best possible Mac experience with native menu bar integration and system-level keyboard shortcuts. Windows and Linux versions are not currently available.

Start Typing Faster Today

Free to try — no credit card required

Download Blurt