voice-api

Try it

Running live on the server this page is served from. Free to use while in beta — a shared key is already filled in, limited to twenty requests a minute per visitor and two at a time.

Text to speak

Voice Speed — 1.00×

Kokoro on CPU. Roughly a second of compute per second of speech — a laptop with Apple silicon is about ten times quicker.

Transcript

Record something, or choose an audio file.

Or upload

Whisper small on CPU, English only. Transcribes at about two thirds of real time, so a clip comes back sooner than it took to record.

🔌

Drop-in OpenAI

Both endpoints are byte-for-byte OpenAI's. Change the base URL and existing SDK code keeps working — no wrapper, no rewrite.

🔒

Nothing leaves the box

Models run on this server, not behind someone else's API. No third party sees your audio or your text, because there is no third party.

🔑

Free while in beta

No signup, no card, no quota beyond a fair-use rate limit. Metered keys and paid plans come later — for now it costs nothing to build against.

API

Base URL https://voice.devops-monk.com · every request needs Authorization: Bearer <key>

POST/v1/audio/speechText to speech

Field	Type	Default	Notes
`model`	string	required	`tts-1` or `kokoro`
`input`	string	required	The text to speak
`voice`	string	af_heart	Any Kokoro English voice
`response_format`	string	mp3	mp3, wav, opus, flac
`speed`	number	1.0	0.25 to 4.0

POST/v1/audio/transcriptionsSpeech to text

Field	Type	Default	Notes
`file`	file	required	wav, mp3, m4a, ogg, flac, webm — max 25 MB
`model`	string	required	`whisper-1` — English, Whisper small Q8_0
`response_format`	string	json	json, text, verbose_json, srt, vtt
`temperature`	number	0	0 to 1. Raise only if output is repetitive

GET/v1/modelsWhat is routable

# speak
curl https://voice.devops-monk.com/v1/audio/speech \
  -H "Authorization: Bearer $KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"tts-1","voice":"af_heart","input":"Hello."}' \
  --output hello.mp3

# transcribe
curl https://voice.devops-monk.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $KEY" \
  -F file=@meeting.m4a -F model=whisper-1

# the official OpenAI SDK works unchanged
from openai import OpenAI
client = OpenAI(base_url="https://voice.devops-monk.com/v1", api_key=KEY)

client.audio.speech.create(model="tts-1", voice="af_heart",
                           input="Hello.").stream_to_file("hello.mp3")

with open("meeting.m4a", "rb") as f:
    print(client.audio.transcriptions.create(model="whisper-1", file=f).text)

Run your own

Four containers, one compose file. No GPU required.

git clone https://github.com/devops-monk/voice-api /opt/voice-api
cd /opt/voice-api && cp .env.example .env
./scripts/deploy.sh

whisper.cpp

Whisper small Q8_0, English. Measured at 0.66× real time on four shared vCPUs.

Kokoro

The same weights the Lector desktop app offers as its quality option.

LiteLLM

Auth, virtual keys, budgets and rate limits. Routing lives in one config file.

This is not a backend for Vox and Lector. Those run entirely on your own machine and always will. This exists for clients that cannot run a 900 MB model — browser extensions, web pages, phones — and it is slower than a laptop. Hosting buys reach, not speed.