Every word.
Down to the millisecond.

Sandchest is the open-source speech-to-text API. Drop audio in the chest and get every word back with real timestamps — from the API your AssemblyAI code already speaks.

Start transcribing See the API

open source · prepaid, billed by the second

Your AssemblyAI code already speaks Sandchest.

Use the official SDK for uploads, polling, word timestamps, and captions. Change the key and the base URL to try the transcription workflow. Additional AssemblyAI features are still being brought across.

2 lines changed · 0 rewritten

transcribe.ts+2 −1

 import { AssemblyAI } from "assemblyai";   const client = new AssemblyAI({-  apiKey: process.env.ASSEMBLYAI_API_KEY,+  apiKey: process.env.SANDCHEST_API_KEY,+  baseUrl: "https://stt-api.sandchest.com", });   const transcript = await client.transcripts.transcribe({   audio: audioBuffer,   disfluencies: true, });

Words you can scrub to.

Whisper, Parakeet, and MMS, with word-level alignment. Each word includes timing, confidence, and punctuation that editors, captions, and search are built on. No placeholder transcripts, no rounded timestamps.

start / end: milliseconds, for every word
confidence: per word, not per file
disfluencies: the ums and uhs, if you ask
language_code: detected automatically or selected explicitly

standup_final_v3.m4a0:04.800

0:00.000

{ "text": "Okay", "start": 0, "end": 280, "confidence": 0.99 }

Our GPUs, or yours.

stt-api.sandchest.com

Hosted

GPU workers start on demand and return results through the job API. The first request after an idle period takes longer while models load. Prepaid credits, metered by the second. No seats, no subscription.

Get an API key →

docker-compose.yml

Self-hosted

Open source, end to end. Your GPU or Apple Silicon, your object storage, your retention window — and you know exactly which model transcribed your audio.

$ git clone https://github.com/CapSoftware/Sandchest
$ docker compose up --build

Read the source →

Go on, open the chest.

Start transcribing