Speaches
An OpenAI-compatible server under MIT for streaming transcription, translation and text to speech, loading Whisper, Kokoro and Piper models on demand, described by its maintainers as Ollama for speech models.
What it does
Speaches "is an OpenAI API-compatible server supporting streaming transcription, translation, and speech generation". Its README's own ambition: "This project aims to be Ollama, but for TTS/STT models." Any tool or SDK that speaks OpenAI's audio API should work against it with a base URL change.
Speech to text is powered by faster-whisper, with transcription streamed back over server-sent events as it happens. Text to speech uses Kokoro and Piper voices. It also offers a realtime API and audio through the chat completions endpoint in three shapes: text in and audio out, audio in and text out, audio in and audio out. Models load on the first request that names them and unload after a period without use.
What it replaces
Speaches replaces Deepgram and AssemblyAI for code already written against OpenAI's transcription endpoint, and replaces ElevenLabs for code written against OpenAI's speech endpoint, because it presents both behind the same API shape on your own hardware.
Where it falls short: the README is short and defers to the website for installation and usage, and it does not state languages, word timestamps, speaker labels or model sizes. The models it loads are third-party and carry their own licences, which it does not list. Its latest GitHub release is a release candidate.
How it runs
Docker or Docker Compose, with GPU and CPU support. Nothing else is stated in the README; the documentation site covers the rest.
Who it fits
A team that wants one container serving both directions of speech behind the API its code already uses. A team that needs speaker labels should pair it with WhisperX, or use OpenWhispr or Vibe on the desktop.
Limits the README states
The streaming transcription demo section reads "TODO". Nothing else is listed.
Sources
- Speaches README and MIT LICENSE: github.com/speaches-ai/speaches, read 2026-09-16.
More speech-to-text engines
OpenAI's general-purpose speech recognition model, released with code and weights under MIT, in six sizes from 39 million to 1.55 billion parameters, with multilingual transcription, translation to English and language detection.
A dependency-free C and C++ port of Whisper under MIT that runs on CPUs, Apple Silicon and phones, with quantised models, an HTTP server and bindings for a dozen languages.
Batched Whisper transcription under BSD-2-Clause with word-level timestamps from forced alignment and speaker labels from pyannote, the closest open-source match to a hosted transcription API's output.
NVIDIA's speech models and toolkit: Apache-2.0 code, and the Parakeet TDT 0.6B v3 model under CC-BY-4.0 with 25 European languages, automatic language detection, punctuation and word timestamps.
An offline speech recognition toolkit under Apache-2.0 with streaming, models of about 50 MB, 20+ languages, a reconfigurable vocabulary and speaker identification, from small boards to clusters.
A runtime under Apache-2.0 from the next-generation Kaldi team that runs speech to text, text to speech, speaker diarization and voice activity detection locally on CPUs, NPUs and phones, with no internet connection.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.