Speaches speech-to-text and voice server for small business
Speaches is a self-hosted speech server. See what businesses use it for and how MILabs sets it up alongside the systems you already run.
Updated 2026-07-07
What Speaches is
Speaches is an open source server that runs speech models on infrastructure the business controls. It offers speech recognition, which turns recorded or live audio into text, and speech synthesis, which turns text into spoken audio. Its interface follows the OpenAI audio API format, so developers who already build against that format can point existing calls at a Speaches instance with few changes. For a small business, the main consideration is where audio goes. Calls, voice notes and meeting recordings can be transcribed on a server that the business or its IT provider manages, so recordings need not pass through a hosted transcription service. That benefit comes with hosting work and with responsibility for consent, retention and access, and someone has to own both. Speaches belongs to the voice and meetings category. Someone still has to decide where transcripts are stored, who reads them and which step follows. Those decisions belong in a plan agreed before anything is switched on. This page covers what Speaches is and how MILabs sets it up. MILabs is based in Sydney, works remotely and runs sessions by video call.
Speaches in plain language
Speech-to-text converts spoken words into written text, and speech synthesis performs the reverse. Many services offering these functions run on a vendor's cloud, so audio leaves the business before it is processed. A self-hosted server changes that arrangement, because the models run on a machine that someone in the business or its IT provider operates. The trade is direct control over where audio sits, in exchange for responsibility for keeping the machine running.
Audio arrives from a phone line, a recording app or an uploaded file. The server returns text, and other systems decide what happens next. Models differ in accuracy, language coverage and the hardware they need, so the choice of model shapes both output quality and running cost. A transcript that misspells a name or street address can cause errors downstream, so output is best checked against real recordings before anyone relies on it.
The OpenAI audio API format is what the server follows, so code written against that style of interface can sometimes be pointed at a local instance. That lowers the cost of trying the tool inside an existing project, although each call path still needs its own testing. Anyone weighing the option should also settle who applies updates, watches for failures and decides on changes.
What businesses use it for
- Transcribing customer calls so staff can search what was said and follow up accurately
- Turning meeting recordings into written notes for team members who could not attend
- Converting voice notes from site staff into job or incident records
- Producing spoken versions of scripted messages for phone menus or appointment reminders
- Creating draft notes from dictated recordings for staff to review before filing
Who it suits
- Businesses with an IT provider or in-house technician who can run a server
- Teams that handle recordings under strict rules about where data is stored
- Developers building voice features who already work with OpenAI-style audio calls
How MILabs sets it up
MILabs starts with the owner or team lead, agreeing which calls, recordings or notes the business wants handled and what must stay on its own infrastructure. Together you choose the hosting arrangement and the models to run, since that choice shapes accuracy and resource use. Next, we connect Speaches to the systems the business already uses, such as its phone platform and CRM, and check that audio reaches the server and text returns where it should. We set permissions so only named accounts can send audio or read transcripts, and we agree in writing who keeps the server patched, monitors it and handles changes. Every step is reviewed with you before the next one begins. MILabs works remotely and runs sessions by video call.
Builds it fits
Industries it fits
Speaches in practice
In a business that runs Speaches, the work usually begins with a defined stream of audio. A phone platform or recording app sends each file to the server, and the server returns a transcript for a second system to store beside its record. Staff read the text where they already work instead of opening a separate tool to replay recordings. Usually the owner and one technical contact change the flow, while everyone else only reads the output.
Audio is often sorted by purpose first. Short voice notes from field staff might feed a job record, while longer meeting recordings feed a notes template that lists decisions and owners. Each stream carries its own rules for who may open the transcript and how long the text is kept, and those rules live in the configuration rather than in anyone's memory. Early on, someone compares sampled transcripts with the original audio, and checks less often once common errors are understood.
Speaches handles only the speech step, so the summary, routing and any reporting sit in separate components that read its output. If the server is unavailable, a sound design keeps the recording and queues it for transcription later, so calls are not lost. Keeping that fallback simple matters more than making the main path clever.
Frequently asked questions
- When is Speaches the wrong choice?
- Speaches suits a business with someone who can run a server, keep it patched and watch its uptime. A hosted speech service with support built into the contract often suits a team that prefers not to run infrastructure. For a single transcription with no privacy concern, a hosted option is the simpler route. Which path fits depends on who will own the hosting and how sensitive the audio is.
- Can my team use Speaches without training?
- Yes for staff who only receive transcripts or spoken output, since those outputs usually arrive in tools they already use. Anyone who sends audio, reads raw output or changes settings needs a short walkthrough with MILabs and a written note of the steps. People who connect Speaches to other systems gain the most from training. Early questions are normal, and a short refresher usually answers them.
- Can MILabs build on top of Speaches?
- Yes, MILabs can build workflows on top of a Speaches server, such as turning call audio into transcripts that feed a CRM record. Each build starts from the business's own data and the systems it already runs, and the team agrees who maintains the server afterwards. Every design is reviewed with the owner before it goes live.
- What does a Speaches setup need from me?
- A Speaches setup needs you to name the calls, recordings or notes to be handled, and to say which systems hold the data around them. Someone must also own the server, apply updates and respond when it fails. Decide who may send audio and who may read transcripts, since setup fixes those permissions. Finally, settle the consent and retention rules for your recordings, because the software does not set those for you.
- What does Speaches not do?
- Speaches does not decide what a transcript means, who should act on it or whether a recording may be kept. Those tasks stay with the business and its team. It handles the speech step only, so summaries, routing and reporting need other components. It also does not run itself, because a server still needs hosting, updates and monitoring from someone accountable.
