engine-selection.md

Engine Selection

RealtimeTTS supports local engines, cloud engines, free-service wrappers, and experimental local voice-cloning stacks. Pick the smallest engine that matches your latency, quality, licensing, and deployment needs.

Status on this page is documentation status, not a guarantee of support level. Each current concrete engine source now has a focused page under docs/engines/.

Fast Choices

Need	Start with
Verify local audio quickly	SystemEngine
Free simple service, no API key	GTTSEngine or EdgeEngine
Commercial cloud API	OpenAIEngine, AzureEngine, ElevenlabsEngine, CartesiaEngine, MiniMaxEngine, CambEngine, or ModelsLabEngine
Word timings	AzureEngine; KokoroEngine for supported English voices
Local neural voice cloning	CoquiEngine, ZipVoiceEngine, OmniVoiceEngine, FasterQwenEngine, LuxTTSEngine, ChatterboxEngine, SoproTTSEngine, PocketTTSEngine, PocketTTSGpuEngine, NeuTTSEngine, or MossTTSEngine
Reliability fallback	Pass a list of compatible engines to TextToAudioStream

Engine Matrix

Engine	Type	Install path observed	Needs	Output/playback notes
SystemEngine	Local system TTS	realtimetts[system]	System voices	PCM through PyAudio.
GTTSEngine	Free service wrapper	realtimetts[gtts]	Network access	Service-backed behavior; no API key.
EdgeEngine	Free service wrapper	realtimetts[edge]	Network access, mpv	Compressed audio path through mpv.
OpenAIEngine	Cloud API	realtimetts[openai]	OPENAI_API_KEY	MP3 uses mpv; PCM uses PyAudio.
AzureEngine	Cloud API	realtimetts[azure]	Constructor key and region	PCM through PyAudio; word timings supported.
ElevenlabsEngine	Cloud API	realtimetts[elevenlabs]	ELEVENLABS_API_KEY, mpv	MPEG audio path through mpv.
CambEngine	Cloud API	realtimetts[camb]	CAMB_API_KEY	API-backed multilingual synthesis.
MiniMaxEngine	Cloud API	realtimetts[minimax]	MINIMAX_API_KEY, mpv	MP3 output, pytest coverage exists.
CartesiaEngine	Cloud API	realtimetts[cartesia]	CARTESIA_API_KEY	Raw PCM formats in source.
ModelsLabEngine	Cloud API	realtimetts[modelslab]	MODELSLAB_API_KEY, mpv	Present in engines package, not root export.
CoquiEngine	Local neural	realtimetts[coqui]	Local model files, likely GPU	Voice cloning; call shutdown() in scripts.
PiperEngine	Local executable	realtimetts[piper] plus Piper binary	Piper binary, model/config files	Very practical for local deployment; uses PIPER_PATH.
StyleTTSEngine	Local neural	realtimetts[styletts] plus StyleTTS2 checkout	StyleTTS2 checkout/model assets	Local voice/reference-audio workflow.
ParlerEngine	Local neural	realtimetts[parler]	Parler stack, GPU recommended	Heavy local model setup.
KokoroEngine	Local neural	realtimetts[kokoro] plus language extras	Optional jp, zh, ko extras	Word timings for supported English voices.
OrpheusEngine	Local/API-style	realtimetts[orpheus]	OpenAI-compatible endpoint or local model stack	LM Studio path documented from Zaphod notes.
FasterQwenEngine	Local neural	realtimetts[qwen]	Faster Qwen3 stack, reference audio	Voice cloning and emotion examples exist.
OmniVoiceEngine	Local neural	realtimetts[omnivoice]	OmniVoice stack, reference audio/text	Multilingual voice cloning.
PocketTTSEngine / PocketTTSGpuEngine	Local lightweight	realtimetts[pockettts], or realtimetts[pockettts-gpu] plus CUDA PyTorch and the pinned GPU fork	pocket-tts package or CUDA fork	CPU-oriented voice cloning path plus optional CUDA fork engine.
NeuTTSEngine	Local neural	realtimetts[neutts], optional neutts-gguf	neutts[llama,onnx] for GGUF streaming	Voice cloning with reference audio.
ZipVoiceEngine	Local neural	realtimetts[zipvoice] plus checkout	ZipVoice checkout/assets or Docker example	Use distill with at least 3 steps for fast quality work.
LuxTTSEngine	Local neural	realtimetts[luxtts]	LuxTTS stack and model assets	Full-waveform local voice cloning.
ChatterboxEngine	Local neural	realtimetts[chatterbox]	chatterbox-tts, reference WAV	Full-waveform prompt-audio workflow.
SoproTTSEngine	Local neural	realtimetts[sopro]	Sopro package, optional HF cache	Native streaming reference-audio workflow.
SopranoEngine	Local neural	realtimetts[soprano]	Soprano package/model assets	Single-voice English streaming.
MossTTSEngine	Local neural	realtimetts[moss] plus runtime assets	MOSS-TTS-Nano assets, CUDA/CUDNN depending on backend	Nano ONNX/torch paths documented; Realtime variant blocked.

Fallback Engines

You can pass a list of engines. RealtimeTTS starts with the first one and can switch when synthesis fails.

from RealtimeTTS import TextToAudioStream, SystemEngine, GTTSEngine


if __name__ == "__main__":
    stream = TextToAudioStream([GTTSEngine(), SystemEngine()])
    stream.feed("Try one engine, then fall back if needed.")
    stream.play()

Fallbacks work best when engines accept compatible text, voices, and output expectations. Avoid mixing engines that require unrelated voice objects unless you handle voice selection per engine.

Licensing

The RealtimeTTS source is MIT licensed, but engine providers, model weights, voice data, and generated audio can carry separate restrictions. Check provider terms before commercial use.