Walkie

Product Introduction

1. Definition: Walkie is a cross-platform AI voice productivity application that combines three distinct technical categories into a single desktop and mobile client: real-time speech-to-text (STT) dictation, AI-powered meeting transcription with speaker diarization, and text-to-speech (TTS) read-aloud. It is distributed by the developer B150 and runs natively on macOS (Apple Silicon), Windows 10+, Linux (x64, Wayland/X11, experimental), and iPhone, with Android listed as forthcoming. The current release is v1.16.2, shipped as a .dmg installer, .exe installer, a one-line curl shell script for Linux, and an App Store build for iOS. Underneath, the application is a system-wide voice layer rather than a browser extension or per-app plug-in — it types directly into whatever text field holds OS focus.

2. Core Value Proposition: Walkie exists to eliminate the keyboard as the primary input bottleneck for knowledge workers. The vendor positions dictation at roughly 220 words per minute against an average typing speed of 45 WPM — approximately a 4x throughput gain — while solving the two objections that historically blocked adoption of voice-to-text software: accuracy cleanup (filler words, grammar, punctuation, formatting) and privacy (audio leaving the machine). Its differentiator is a dual-mode architecture: an On-Device Mode that performs unlimited, fully offline, no-account dictation with 0 ms network latency, and an opt-in Cloud Mode that applies LLM-based polish, filler removal, translation, and app-aware formatting. Primary keyword clusters the product targets include AI dictation software, offline speech-to-text, meeting transcription with speaker labels, and natural-sounding text-to-speech.

Main Features

1. Dual-Mode Transcription Engine (On-Device + Cloud): Walkie does not depend on a single inference path. On-Device Mode runs open-weight speech recognition models locally — the supported model families are OpenAI Whisper (Small, Medium, and Turbo variants), NVIDIA Parakeet (English-only, tuned for higher accuracy), and Moonshine — with 7+ models available for download and hot-swapping. Local inference is hardware-accelerated through Metal on macOS, Vulkan on Windows and Linux, and OpenBLAS for CPU-bound workloads. Because the models are resident, on-device transcription returns text with no network round trip, works with Wi-Fi disabled, and requires no account. Cloud Mode routes a short audio clip to a provider for transcription plus formatting in a single pass; the vendor states clips are discarded after the text is returned, that providers are contractually bound not to train on customer audio or transcripts, and that Cloud Mode is an explicit opt-in rather than a silent upload. Cloud Mode also powers voice translation — speaking "translate to Spanish" before a message outputs the result in the target language. Language coverage is 100+ languages via the Whisper family.

2. Meeting Transcription with Speaker Attribution: Rather than joining a call as a visible bot participant, Walkie captures system audio directly from the host computer, recording both the local microphone and the remote participants' audio while labeling who said what. It generates running notes, summaries, action items, and a permanent transcript library as the meeting proceeds, with no meeting-bot invite appearing in the participant list. Meeting allowances scale by tier: a 3-meeting trial is included in the free plan, Pro covers 10 meetings per month up to 8 recorded hours, Work covers 100 meetings per month up to 40 hours, and Team covers 150 meetings per month up to 75 hours. An experimental MCP server exposes the meeting library to external AI agents such as Claude and Cursor, and 10+ experimental app connectors (Asana, Linear, Slack, GitHub, and others) push meeting outputs into project management and communication tools directly. Recorded hours are counted identically whether captured on-device or via the cloud.

3. Read Aloud (950+ Voice Text-to-Speech): Walkie's TTS subsystem converts PDFs, highlighted text, or right-clicked selections into speech using a library of 950+ natural voices, with synchronized word-level highlighting that tracks the spoken position in the source text. Triggering is fully system-integrated: drop in a document, press a global shortcut on any highlighted passage, or use the context menu. Free accounts receive 10 minutes per month of AI voice output; Pro and above receive 2 hours per month, and Work and Team receive 5 hours per month.

4. Context-Aware Formatting and AI Auto Edits: In Cloud Mode, transcription and formatting happen in one inference step. The pipeline strips filler words (the vendor cites a 94% filler-removal rate), corrects grammar, applies punctuation, and restructures rambling speech into readable prose. App-aware formatting adapts the same spoken words to their destination — emails receive a greeting, body, and sign-off; chat and messaging apps stay casual on a single line; code editors and terminals remain literal and unformatted. A separate Text Style control offers Formal, Casual, and Very Casual registers. The vendor cites an edit accuracy figure of 97.8% and a formatting latency of 12 ms.

5. Personalization Layer: Smart Dictionary, Live Dictionary, and Snippets: The Smart Dictionary learns user vocabulary automatically — correcting a mistranscribed word prompts a quiet top-right toast offering a one-tap "Add to dictionary" action, so nothing is learned without explicit consent. Industry terms, proper names, and unusual spellings can also be entered manually. Snippets map trigger phrases to expanded text (for example, a saved "my email" trigger expanding to a full address), eliminating repeated spelling of static content. Together these address the domain-vocabulary failure mode that makes generic dictation unusable for developers, clinicians, and legal professionals.

6. Voice Commands, Ask Agent, and Experimental Extensions: Walkie accepts spoken system commands — "open cursor," "open gmail," "open notes" — that launch applications, open URLs, and control workflow without touching the keyboard. The experimental Ask Agent accepts a held-shortcut question and speaks the answer back, with live weather, news headlines, web search, and actions such as sending email. Sentence Rewrite (macOS-only, experimental) lets a user drop the cursor mid-sentence and dictate replacement text; Walkie rewrites the entire sentence rather than splicing in the new words, avoiding grammatical mash-ups. Smart Commands for chaining actions and a custom prompt library are listed as upcoming.

Problems Solved

1. Pain Point — Slow Manual Input: Typing caps knowledge-worker output at roughly 45 WPM and breaks flow during idea capture. Walkie removes that ceiling by converting speech to finished text at approximately 220 WPM and inserting it directly into the focused application, so no copy-paste or secondary window is required.

2. Pain Point — Unusable Raw Transcription: Conventional dictation tools output verbatim speech littered with "um," "uh," false starts, and missing punctuation, forcing users to spend more time editing than they saved. Walkie solves this with automatic filler detection and removal, grammar correction, punctuation insertion, and destination-specific formatting — a spam of text becomes a structured email or a single-line Slack message.

3. Pain Point — Cloud Privacy Exposure: Cloud-only voice tools transmit audio to remote servers, require persistent internet connectivity, may retain data under undefined retention policies, and in some cases use audio for model training. Walkie's On-Device Mode keeps audio on the machine, works offline, requires no account, and the vendor advertises HIPAA-ready operation with Zero Data Retention and a contractual no-training guarantee from cloud providers.

4. Pain Point — Meeting Notes Overhead: Manual note-taking during calls splits attention and produces incomplete records; bot-joined transcription tools visibly announce themselves in the participant list and alter meeting dynamics. Walkie records from the local machine without a bot, produces labeled who-said-what transcripts, and retains them indefinitely in a searchable meeting library.

5. Pain Point — Domain Vocabulary Errors: Off-the-shelf speech engines mangle product names, technical jargon, and proper nouns. The Smart Dictionary and Live Dictionary mechanisms let corrections persist automatically, meaning accuracy improves with use instead of resetting every session.

6. Target Audience: Software engineers and developers dictating commit messages, code comments, CLI commands, and AI coding prompts into VS Code, Cursor, Terminal, Warp, Claude Code, and GitHub; technical writers and documentation engineers authoring in Notion, Obsidian, Google Docs, and Evernote; product managers and project leads managing tickets and specs in Linear, Jira, Trello, ClickUp, and Monday; sales, support, and marketing staff operating in Slack, Gmail, LinkedIn, WhatsApp, and X; compliance-sensitive professionals in healthcare and legal who require HIPAA-ready, offline-capable transcription; accessibility users who rely on speech-to-text and text-to-speech as primary interaction modalities; and multilingual users leveraging 100+ language support and voice translation.

7. Use Cases: Dictating a polished client email into Gmail without leaving the inbox; writing commit messages and pull-request descriptions inside VS Code or Cursor; capturing meeting notes and action items during a Zoom or Teams call with speaker attribution; converting a PDF research paper into audio for commute listening via Read Aloud; issuing voice commands to launch applications and URLs during a heads-down coding session; drafting long-form articles or newsletters in Substack and Medium at speaking speed; transcribing client interviews offline in an air-gapped or low-connectivity environment; asking the Ask Agent for live weather, news, or web results hands-free; building a personal dictionary of product names and technical terms that persists across all applications.

Unique Advantages

1. Differentiation: Most dictation products in this category are cloud-only, subscription-gated, and English-centric. Walkie's structural difference is mode choice within one binary: an unlimited free offline engine for privacy-critical or connectivity-free work, and a cloud path for speed and LLM polish. That means free-tier users are not limited by word count on-device (no trials, no credit card), whereas competing free tiers typically meter total transcription. It is also multi-surface rather than single-surface — the same application handles dictation, meeting capture, and read-aloud, replacing three separate tools. Platform coverage extends further than most competitors, which typically ship macOS and Windows only: Walkie ships a Linux build (x64, Wayland/X11) with a one-command installer and an Omarchy bar widget for Arch, plus an iPhone app, with Android pending. Pricing undercuts the category at $6/month annual for Pro, $14/month for Work, and $25/month for Team, with the free tier serving as a functional product rather than a demo.

2. Key Innovation: The distinguishing architectural decision is treating device-local inference as the default rather than a premium upsell. Walkie ships 7+ downloadable speech models — including Whisper Small, Medium, and Turbo, NVIDIA Parakeet for English accuracy, and Moonshine — and accelerates them through Metal, Vulkan, and OpenBLAS so that offline transcription is fast enough for real-time dictation. Layered on top is app-aware formatting that changes output shape based on the destination text field (greeting/body/sign-off for email, single-line casual for chat, literal for terminals), and a consent-gated Live Dictionary that converts user corrections into permanent vocabulary. The combination means the system is simultaneously more private, more accurate over time, and more contextually appropriate than a generic speech-to-text pipeline.

Frequently Asked Questions (FAQ)

1. Is Walkie free, and what does the free plan actually include? Yes — Walkie Free is $0 forever with no trials, no word limits, and no credit card required. It includes unlimited On-Device Mode dictation, 2,000 words per week of Cloud Mode transcription and formatting, 10 minutes per month of AI Read Aloud voices, 7+ downloadable on-device models, AI text styling, 100+ languages, system-wide operation in every app, privacy mode with Zero Data Retention, and a 3-meeting trial of desktop meeting notes. Paid tiers add unlimited Cloud Mode, more meeting hours, and expanded voice minutes: Pro at $6/month billed annually ($72), Work at $14/month ($168), and Team at $25/month ($300).

2. Does Walkie work offline, and is my audio really kept on-device? On-Device Mode works fully offline once the speech model has been downloaded — Wi-Fi can be disabled and transcription continues with 0 ms network latency, and no account is required. Audio never leaves the computer in this mode. Cloud Mode is a separate, explicitly opt-in path that sends a short audio clip to a provider for transcription and formatting; Walkie states the clip is discarded after the text returns, providers are bound not to train on user audio or transcripts, and the product is HIPAA-ready with Zero Data Retention. Users can switch between modes at any time.

3. Which platforms and apps does Walkie support? Walkie runs on macOS (Apple Silicon) as a .dmg, Windows 10+ as a .exe installer, Linux (x64, Wayland/X11, experimental) via a one-line curl -fsSL https://trywalkie.com/install.sh | sh install that keeps itself updated, and iPhone via the App Store; Android is listed as coming soon. Because it operates system-wide and types directly into the focused text field, it works in Slack, Gmail, Notion, VS Code, Terminal, Apple Notes, Messages, Google Docs, Cursor, Claude Code, Antigravity, and any other application with a text input.

4. What speech recognition models and languages does Walkie use? Walkie supports OpenAI Whisper (Small, Medium, Turbo), NVIDIA Parakeet, and Moonshine, with 7+ models available for download and manual switching depending on whether throughput, accuracy, or English-only precision is the priority. Whisper-based models cover 100+ languages including English, Simplified and Traditional Chinese, Cantonese, German, Spanish, Russian, Korean, French, Japanese, Portuguese, Turkish, Polish, Arabic, Hindi, Vietnamese, Hebrew, Ukrainian, Greek, and dozens more. Parakeet is English-only but tuned for higher accuracy on English audio.

5. How does Walkie remove filler words and fix grammar automatically? In Cloud Mode, Walkie performs transcription and formatting in a single inference pass rather than two sequential steps. That pass detects and strips filler words such as "um" and "uh" (vendor-cited 94% filler removal), corrects grammar, applies punctuation and capitalization, and structures the result according to the active application — email formatting with greeting and sign-off, casual single-line formatting for chat apps, and literal unformatted output for code editors and terminals. The vendor reports 97.8% edit accuracy and 12 ms formatting latency. Free accounts receive 2,000 Cloud Mode words per week; Pro and above remove that cap entirely.

6. Can Walkie transcribe meetings without a bot joining the call? Yes. Walkie records system audio directly from the host computer, capturing both the user's microphone and incoming participant audio with speaker labels, so no bot appears in the meeting participant list and no meeting invite is required. It generates summaries, action items, and a who-said-what transcript library that is retained permanently on the Work and Team tiers. Meeting limits are 3 meetings (free trial), 10 meetings/8 hours (Pro), 100 meetings/40 hours (Work), and 150 meetings/75 hours (Team), with an experimental MCP server connecting the meeting library to Claude, Cursor, and other AI agents.