NewSochen Dictate for Windows

Talk to your coding agents.Sochen types it.

Hold a key anywhere on Windows and speak: English, Hindi, Hinglish, or eight more Indian languages. Let go, and clean text, or a sharp prompt for Claude Code, Codex or Cursor, is pasted where your cursor is.

Download for WindowsTry it in your browser

Free for 2,000 words a week. Runs on the AI models you choose; Sochen’s defaults are picked by our open benchmark.

Works where you type

  • Claude Code✦
  • Codex✦
  • Cursor✦
  • VS Code✦
  • Windows Terminal✦
  • Slack✦
  • Gmail✦
  • Notion✦
  • WhatsApp Web✦
  • Word✦
  • Chrome✦

Why Sochen

A prompt, not a transcript.

Dictation tools give you your words back. Sochen gives your coding agent what it needs: names in backticks, the fillers gone, the structure your speech had and no more.

“camel case get user name in lib slash auth dot ts”

`getUserName` in `lib/auth.ts`

Claude Code in front · prompt style

Prompts for coding agents

Say “camel case get user name” and get `getUserName`. Paths, commands, flags and env vars in backticks. It knows when Claude Code, Codex, Cursor or VS Code has the cursor, and writes for them.

yaar is function ko refactor karo, null return kare, aur ek test add karo

Refactor this function to return null, and add a test.

Or keep it as spoken, or in Devanagari

Hinglish in, English out

Speak the way you talk. Get the prompt in English, keep it as spoken, or have Hindi in Devanagari. Priced for India: ₹249 a month.

in memorytext

No audio kept. No transcripts kept.

Nothing kept

We never store your audio or your transcripts: each piece is turned into text in memory and dropped. The model provider doesn't train on it (paid API). Your history lives on your device only.

Key downKey up
  • Heard while you were still talking
  • Nothing waits for a connection to open

Never lose a dictation

Recording starts the instant you press the key, before anything connects. Your speech goes up in pieces while you talk, so most of the listening is done by the time you let go.

POST /api/v1/dictate/sessions

Authorization: Bearer sk_sochen_…

{

"id": "dict_…", "clientToken": "…",

"maxSeconds": 600

}

$0.60 an hour of audio, cleanup included

A real API

Sessions, uploads, finish and clean, with keys for your server and per-session tokens for a browser. $0.60 an hour of audio, cleanup included.

How it works

Hold. Speak. Let go.

1

Hold the key

Anywhere on Windows. Recording starts the instant you press, before anything connects.

2

Say it your way

Ums, false starts, Hinglish, “camel case get user name”: all understood.

3

Let go

Clean text, or a prompt for your coding agent, is pasted where your cursor is.

Understands

  • English✦
  • Hindi✦
  • Hinglish✦
  • Bengali✦
  • Tamil✦
  • Telugu✦
  • Marathi✦
  • Gujarati✦
  • Kannada✦
  • Malayalam✦
  • Punjabi✦

Benchmarks

Accuracy and cost, measured, not claimed.

Every listener Sochen can use, run on the public leaderboard’s own English test data, plus Hindi, and the whole engine on spoken coding prompts. The default listener, Grok STT at about $0.10 an hour, beats Whisper large-v3’s published scores on four of six sets and ties the other two.

95.9%

Code words exactly right

Identifiers and file paths in 60 spoken coding prompts, with your dictionary on.

96.6%

Faithful to what you said

Prompts judged to keep your meaning, nothing added or dropped.

0 of 16

Instructions answered

Of 16 dictations that sound like orders to an AI ("ignore the rules above…"): written down, never obeyed.

100%

Silence stays silent

Silence, room noise and music produced no invented text on the default listener.

Listening accuracy in English: word error rate, lower is better

ModelLibriSpeech cleanLibriSpeech otherVoxPopuli (cleaned)Earnings-22AMI meetings (cleaned)Monsoon Indian EnglishCost an hour
Grok STTSochen default1.63.61.27.89.43.3$0.10
Gemini 3.5 Transcribe1.13.21.39.17.63.0$0.18
GPT Transcribe (direct)1.52.73.09.38.03.4$0.28
GPT-4o mini Transcribe1.03.11.88.711.04.0$0.10
Whisper (large-v2)1.63.42.19.920.33.1$0.36
whisper-large-v3 (published, full sets)1.63.54.511.613.64.0–
scribe_v2 (published, full sets)1.12.31.48.98.23.3–
universal-3-5-pro (published, full sets)1.12.11.910.010.63.5–

Hindi (FLEURS, word error rate)

  • Gemini 3.8 Flash8.2
  • Gemini 3.5 Transcribe8.4
  • Gemini 3.5 Flash-Lite9.5
  • GPT-4o mini Transcribe12.1

Hindi goes to the most accurate listener here; Hinglish goes to Grok STT, which ranked first on the MUCS 2021 test and answers fastest.

How we measured

English: 50 clips from each of the Open ASR Leaderboard’s own test sets (40 for Indian English), scored with its normalizer. Published rows are the leaderboard’s full-set scores, shown for scale; differences under about a point on one set are within noise. Coding prompts and the instruction and noise clips are synthetic. Run 2026-09-27 from India, for $3.21 in model calls.

Compared

Sochen Dictate and Wispr Flow.

The price is the one thing we can check for you. Everything else here is about Sochen; try both and keep the one that writes what you meant.

Sochen Dictate

$10a month

₹249 in India

  • Prompts for coding agents: names, paths and commands in backticks
  • Hinglish in, English out; eleven languages in
  • Audio and transcripts never stored; each piece is turned into text in memory and dropped
  • A public API at $0.60 an hour of audio, cleanup included
  • Free for 2,000 words a week

Wispr Flow

$15a month

or $12 a month billed annually

A good dictation tool, and the one many people compare us with. The prices are from wisprflow.ai, September 2026; we don’t make claims about its features here.

Pricing

$10 a month. ₹249 in India.

Start free, no card needed. Plus and Pro include unlimited dictation too, with the most accurate second pass.

Free

$0forever

2,000 words a week

  • Windows app, browser composer and playground
  • Recordings up to 2 minutes
  • Every language, every style
  • 150 agent credits a week, as on Free
Start free

Dictate

$10a month

250,000 words a month, no weekly cap

  • Recordings up to 10 minutes
  • Every language, every style, every app
  • The Free agent allowance stays: 150 credits a week
  • Cancel any time
Start free, then upgrade

The API is charged in credits on every plan: $0.60 an hour of audio, cleanup included. Read the docs

Questions

Does Sochen keep my audio or what I said?+

No. Your speech goes to Sochen's server in short pieces and is passed to the AI model you chose. We never store your audio or your transcripts: each piece is turned into text in memory and dropped. The model provider doesn't train on it (paid API). Your dictation history is kept only on your own device, and you can clear it.

Which languages does it understand?+

English, Hindi and Hinglish, plus Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam and Punjabi. You choose how Hinglish comes out: in English, as spoken, or in Devanagari.

Which apps does it work in?+

Any Windows app with a text cursor. It recognises coding agents and editors, such as Claude Code, Codex, Cursor, VS Code and Windows Terminal, and writes prompts for them; chat gets a lighter style, email a more formal one, and you can change all of that per kind of app.

Is there a Mac version?+

Coming soon. The engine, your settings and your dictionary are the same, so nothing needs setting up twice. Until then, dictation also works in your browser: the mic in Sochen's composer, and the playground on the Dictation page.

Which AI models does it use?+

Runs on the AI models you choose: one listens, one writes the words up. Sochen’s defaults are picked by our open benchmark, and you can pick others from the Dictation page.

Is there an API?+

Yes. Sessions, uploads, finish and clean, with keys for your server and per-session tokens for a browser. $0.60 an hour of audio, cleanup included, charged in credits on any plan. Read the docs.

Just say it.

बोलें means “speak”. Hold the key, say what you want done, and get on with it.

Download for WindowsTry it in your browser