Transcription server

The transcription server is a third dictation engine, next to the system recognizer and offline Whisper. The keyboard records what you say and sends the recording to a server you choose. The server sends back the text. The heavy work runs on a computer on your desk or in your rack, not on your phone’s battery, and you choose the model.

It works in every edition, Lite included. There is nothing to download on the phone.

Setting it up

WM Keyboard›Voice typing

1. Under Engine, set Recognition engine to Transcription server.
2. In the Transcription server group that appears, fill in Server address. For a server on your own network it looks like http://192.168.1.10:8000/v1.
3. Fill in Model if your server needs one (see the recipes below). Fill in API key if it asks for one.
4. Press Test on Test the server. The keyboard sends one second of silence and says whether the server answered. If it fails, it shows the same message a failed dictation would show.

Dictation then works the way it does with offline Whisper. Speak, then press the microphone when you are done (or let the 30-second clip fill). The phrase lands a moment later, with spoken punctuation, spacing and undo applied as usual. Keep listening chains the next phrase the same way.

Send the keyboard language is on by default. It tells the server the language of your active layout, which is faster and more accurate than letting it guess from a short clip. Turn it off if you dictate in two languages on one layout and want the server to detect which one you are speaking.

The address

Type the server’s API root, up to and including the version segment, the same way the AI tool’s Other service asks for it. The keyboard adds /audio/transcriptions itself.

You type	The keyboard sends to
http://192.168.1.10:8000/v1	http://192.168.1.10:8000/v1/audio/transcriptions
http://192.168.1.10:8000	http://192.168.1.10:8000/v1/audio/transcriptions (a bare address gets /v1)
https://api.groq.com/openai/v1	https://api.groq.com/openai/v1/audio/transcriptions
http://192.168.1.10:8080/inference	used as it is (whisper.cpp’s own endpoint)

An address that already ends in /transcriptions or /inference is used as typed.

Running a server

The keyboard does not use a protocol of its own. It speaks OpenAI’s audio transcription API, which most speech-to-text servers already support. Any of these works.

speaches (faster-whisper)

speaches runs faster-whisper behind an OpenAI-compatible API. On a machine with Docker:

docker run --detach --publish 8000:8000 --name speaches \
  --volume hf-hub-cache:/home/ubuntu/.cache/huggingface/hub \
  ghcr.io/speaches-ai/speaches:latest-cpu

Use the latest-cuda image with --gpus=all on a machine with an NVIDIA GPU. speaches only runs models you have downloaded, so fetch one once:

curl -X POST http://localhost:8000/v1/models/Systran/faster-whisper-small

Then set:

• Server address: http://<computer's address>:8000/v1
• Model: the id you downloaded, here Systran/faster-whisper-small. For English only, Systran/faster-distil-whisper-small.en is faster.

whisper.cpp

whisper.cpp ships whisper-server, a single binary with no Python. Download a model and start it on all interfaces:

whisper-server -m models/ggml-base.bin -l auto --host 0.0.0.0 --port 8080 \
  --inference-path /v1/audio/transcriptions

With --inference-path as above, set Server address to http://<computer's address>:8080/v1. Without it, point the keyboard at http://<computer's address>:8080/inference. Leave Model empty: whisper-server uses the model it was started with. -l auto matters when Send the keyboard language is off. Without it whisper-server assumes English.

Cloud services

The same engine reaches the hosted APIs, if you would rather pay per minute than run a machine:

Service	Server address	Model
Groq	https://api.groq.com/openai/v1	whisper-large-v3-turbo
OpenAI	https://api.openai.com/v1	whisper-1 or gpt-4o-mini-transcribe

Both need an API key. LocalAI and other OpenAI-compatible gateways work too.

The protocol

This is everything the keyboard sends and expects, so you can build or proxy a server of your own.

Request. POST <endpoint> with Content-Type: multipart/form-data, and these parts:

Part	Sent	Contents
file	always	the clip, speech.wav, audio/wav: 16-bit PCM, 16 kHz, mono, up to 30 seconds (under 1 MB)
model	when Model is filled in	the text of the field
language	when Send the keyboard language is on and Whisper knows the language	an ISO-639-1 code such as en, bn or de
prompt	when there are words to listen for, or Prompt is filled in	your words to listen for, then the text of Prompt. See Words and the prompt
response_format	always	json

An Authorization: Bearer <key> header is added when API key is filled in. Clips shorter than 0.3 seconds, and clips with nothing louder than a quiet room in them, are never sent. See A silent clip types nothing.

Reply. HTTP 200 with {"text": "the words"}. Other fields are ignored. A plain-text body is accepted too, and is used as the transcript.

Errors. Any status outside 200–299 fails the phrase. When the body is {"error": {"message": "…"}}, that message is what the voice panel shows. Otherwise the panel shows the keyboard’s own wording for the status (401 and 403 read as a rejected key, 429 as a rate limit).

Timeouts are 10 seconds to connect and 60 seconds for the reply.

Words and the prompt

A server gets the words to listen for inside OpenAI’s prompt field, followed by whatever you type into Prompt. The model reads the prompt as the text that came before your clip, so a name in it is a name the model has just “seen” spelled your way. The Prompt field is also where to set a style, such as “Casual chat between friends.” or a sentence in the spelling you want.

Whisper models, which is what most servers run, read only the last 224 tokens of the prompt, and drop the rest without saying so. That’s a limit of the model: Whisper’s decoder has room for 448 tokens and gives half of them to the prompt. 224 tokens is roughly 150 short English words, fewer for long names or for scripts that take several tokens a character, so it can’t hold a whole dictionary. Newer models such as OpenAI’s gpt-4o-transcribe read far more.

So the keyboard builds the prompt for that limit. It sends at most 150 words, and puts them in order of importance with the most important at the end, where a Whisper model keeps them: your Prompt text last, then your own list, then the words you added by hand and the personal dictionary, and the learned words at the front. When a Whisper server has to cut, it cuts the learned words first. The whole prompt is capped at 4,000 characters.

Privacy

With this engine, the recording of each phrase leaves your phone. It goes to the address you typed and nowhere else, and only after you start dictating. Nothing is sent in the background. The API key is stored with your other settings and is left out of backups unless you choose to include API keys. See Network policy.

An http:// address sends the audio unencrypted. That is fine on your own network. Use https:// for anything reached over the internet.

With Listen for my words on, the prompt carries up to 150 of your own words, as described in Words and the prompt. Turn it off, and leave More words to listen for and Prompt empty, and no prompt is sent at all.

Voice typing stays off in password fields, whichever engine is chosen.

Details & edge cases

• Mobile data. Data saver has a Transcription server row. It is allowed by default. Set it to Ask each time and the first phrase on mobile data is held with a note. Pressing the microphone again sends it and allows the rest of that session. Turn off stops the engine on mobile data.
• No address yet. Choosing the engine without filling in the address makes the voice panel say so, with a shortcut to these settings and a button that switches back to the system recognizer.
• Server unreachable. The panel shows why: no connection, a timeout, or a server that refused the connection. Nothing is typed and the rest of your text is untouched. Press the microphone to try again.
• No live words. Like offline Whisper, the server gets the whole phrase at once, so nothing appears in the field while you speak.