S2S

An high-performance, lightweight API server written in Rust that provides local, privacy-conscious Speech-to-Text (STT) and Text-to-Speech (TTS) capabilities. By leveraging the sherpa-onnx framework, S2S offers state-of-the-art inference with minimal latency, requiring no external cloud dependencies.

The project aims to provide a drop-in local alternative for speech processing, featuring an API structure inspired by industry standards.

Key Features

• High Performance: Built with Rust and ONNX Runtime for efficient CPU/GPU utilization.
• Privacy-First: All processing is done locally on your hardware.
• Automated Model Management: Built-in bootstrap logic to download and configure necessary models (Kokoro and Parakeet) automatically.
• Broad STT Language Support: Supports 25+ languages including English, Spanish, German, French, Russian, and many more.
• Flexible TTS: Integration with the Kokoro model, supporting over 50 distinct voices across 9 major languages.
• Robust STT: Powered by the NVIDIA Parakeet TDT model for highly accurate transcriptions.

---

Getting Started

Installation

Download the latest executable for your platform from the Releases page.

Running the Server

Simply run the executable to start the server. On the first run, the application will ask for permission to download the required ONNX models (~1GB total).

./s2s

Alternatively, you can skip the prompts by using the --download flag:

./s2s --download

Configuration Options

The server can be customized via command-line arguments:

Argument	Description	Default
--host	The IP address to bind the server to	127.0.0.1
--port	The port to listen on	35248
--kokoro	Path to the Kokoro TTS model directory	./models/kokoro...
--parakeet	Path to the Parakeet STT model directory	./models/sherpa...
--threads	Number of threads for inference	4
--download	Automatically download missing models	false

---

API Reference

1. Speech-to-Text (STT)

Endpoint: POST /v1/audio/transcriptions

Transcribe an audio file to text. The endpoint expects a multipart/form-data request containing a WAV file. The model automatically detects the language from the supported list.

Request:

curl http://127.0.0.1:35248/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F "file=@audio.wav"

Response:

{
  "text": "Hello world, this is a local transcription."
}

2. Text-to-Speech (TTS)

Endpoint: POST /v1/audio/speech

Synthesize text into high-quality audio.

Request Body:

Field	Type	Description
input	String	The text to be synthesized
voice	String	(Optional) The voice ID (Default: af_alloy)

Example:

curl http://127.0.0.1:35248/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Hello, I am a locally hosted voice.",
    "voice": "af_bella"
  }' --output output.wav

---

Language & Voice Support

Speech-to-Text (STT) Languages

S2S supports transcription for the following languages:

Bulgarian (bg)	Croatian (hr)	Czech (cs)	Danish (da)	Dutch (nl)
English (en)	Estonian (et)	Finnish (fi)	French (fr)	German (de)
Greek (el)	Hungarian (hu)	Italian (it)	Latvian (lv)	Lithuanian (lt)
Maltese (mt)	Polish (pl)	Portuguese (pt)	Romanian (ro)	Slovak (sk)
Slovenian (sl)	Spanish (es)	Swedish (sv)	Russian (ru)	Ukrainian (uk)

Text-to-Speech (TTS) Voices

For TTS, the language is determined automatically based on the prefix of the selected voice.

Language	Voice Prefix	Examples
English (US)	af_, am_	af_alloy, af_sky, am_adam, am_echo
English (UK)	bf_, bm_	bf_alice, bm_daniel
Spanish	ef_, em_	ef_dora, em_alex
French	ff_	ff_siwis
Hindi	hf_, hm_	hf_alpha, hm_psi
Italian	if_, im_	if_sara, im_nicola
Japanese	jf_, jm_	jf_alpha, jm_kumo
Portuguese	pf_, pm_	pf_dora, pm_santa
Chinese	zf_, zm_	zf_xiaobei, zm_yunxi

---

Requirements

• Operating System: Linux, macOS, or Windows.
• Audio Format: For STT, input must be in WAV format (16kHz mono recommended).
• Disk Space: Approximately 1.5GB for models and dependencies.

Acknowledgments

• sherpa-onnx for the underlying inference engine.
• Kokoro for the high-quality TTS weights.
• NVIDIA for the Parakeet TDT ASR models.