Models

Two sizes. One voice.

Open weights. Same training data.

On this phoneOffline

Mama — voice note

0:12

Customer call

1:48

Market price

0:31

Mini

Fast and small. Runs on a laptop CPU or a mid-range phone.

Batch jobTok234 Large

[—]hours transcribed

calls_2026-10.zip

Done

field_mile12/

Done

support_queue/

Running

Large

Our most accurate model. For servers and batch transcription.

Compare

Choose by where it runs

MiniLarge

Runs onPhone, laptop, offlineGPU server

Best forApps and voice notesBatch transcription

Word error rate[—]%[—]%

Parameters[—]M[—]B

Quickstart

from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
    model="tala-labs/tok234-mini")
asr("voice-note.wav")["text"]

Limits

01Accuracy drops with heavy background noise and overlapping speakers. [quantify]
02Code-switching into Yoruba, Igbo or Hausa is only partly covered. [quantify]
03Not tested for medical, legal or emergency decisions. Keep a person in the loop.

Say it your way. Try Tok234 in your browser, or download the weights.