Skip to main content
RCLI supports 6 TTS voices across 4 architectures, ranging from lightweight Piper voices to high-quality Kokoro models with 103 speakers.

​
Quick Comparison

VoiceArchitectureSizeSpeakersQualityLanguagesLicense
Piper Lessac ⭐VITS60 MB1GoodEnglishMIT
Piper AmyVITS60 MB1GoodEnglishMIT
Matcha LJSpeechMatcha100 MB1GreatEnglishMIT
KittenTTS NanoKitten90 MB8GreatEnglishApache 2.0
Kokoro English v0.19 🏆Kokoro310 MB11ExcellentEnglishApache 2.0
Kokoro Multi-lang v1.1Kokoro500 MB103ExcellentChinese + EnglishApache 2.0
⭐ = Default (ships with rcli setup)
🏆 = Recommended upgrade for best quality

​
Voice Details

​
Piper Lessac (Default)

Recommended for most users. Fast synthesis, clear English, low latency.

​
Specifications

  • Provider: Rhasspy Piper
  • Architecture: VITS (Variational Inference TTS)
  • Size: ~60 MB
  • Speakers: 1 (male voice)
  • Quality: Good
  • Languages: English (US)
  • Latency: ~150.6 ms average on M3 Max
  • License: MIT
  • Download: rcli setup (default)

​
Model Files

Stored in ~/Library/RCLI/models/piper-voice/: 
piper-voice/
├── en_US-lessac-medium.onnx
├── en_US-lessac-medium.onnx.json
└── tokens.txt

​
Key Features

  • Fastest TTS synthesis (~150ms latency)
  • Clear, neutral male voice
  • Optimized for command responses
  • Minimal disk footprint (60 MB)
  • Double-buffered playback (next sentence synthesizes while current plays)

​
When to Use

  • Default choice for most users
  • Fast response times required
  • Limited disk space
  • English-only use cases

​
Piper Amy

​
Specifications

  • Provider: Rhasspy Piper
  • Architecture: VITS
  • Size: ~60 MB
  • Speakers: 1 (female voice)
  • Quality: Good
  • Languages: English (US)
  • License: MIT
  • Download: rcli voices

​
Description

Warm female alternative to Piper Lessac. Same speed and quality, different tone.

​
KittenTTS Nano

​
Specifications

  • Provider: KittenML
  • Architecture: Kitten (custom architecture)
  • Size: ~90 MB
  • Speakers: 8 (4 male, 4 female)
  • Quality: Great
  • Languages: English
  • License: Apache 2.0
  • Download: rcli voices

​
Model Files

Stored in ~/Library/RCLI/models/kitten-nano-en-v0_1-fp16/: 
kitten-nano-en-v0_1-fp16/
├── model.fp16.onnx
├── tokens.txt
└── voices.bin

​
Key Features

  • 8 distinct voices — Choose from 4 male and 4 female speakers
  • Lightweight (90 MB for all 8 voices)
  • Good quality for size
  • Voice ID selectable via config

​
Speaker Selection

Switch speakers in the TUI or via config: 
# ~/.Library/RCLI/config
tts_model=kitten-nano
tts_speaker=3  # Speaker IDs: 0-7

​
Matcha LJSpeech

​
Specifications

  • Provider: Matcha-TTS
  • Architecture: Matcha (flow-matching)
  • Size: ~100 MB (model + HiFi-GAN vocoder)
  • Speakers: 1 (female voice)
  • Quality: Great
  • Languages: English
  • License: MIT
  • Download: rcli voices

​
Model Files

Stored in ~/Library/RCLI/models/matcha-icefall-en_US-ljspeech/: 
matcha-icefall-en_US-ljspeech/
├── model-steps-3.onnx
├── hifigan_v2.onnx
└── tokens.txt

​
Description

Fast synthesis with clear female voice. Uses HiFi-GAN vocoder for high-quality audio generation.
Best English quality. 11 natural-sounding voices with excellent prosody.

​
Specifications

  • Provider: Hexgrad
  • Architecture: Kokoro (82M parameters)
  • Size: ~310 MB
  • Speakers: 11 (various voices)
  • Quality: Excellent
  • Languages: English
  • License: Apache 2.0
  • Download: rcli voices

​
Model Files

Stored in ~/Library/RCLI/models/kokoro-en-v0_19/: 
kokoro-en-v0_19/
├── model.onnx
├── tokens.txt
└── voices.bin

​
Key Features

  • 11 unique voices — Wide variety of tones and styles
  • Best prosody and naturalness
  • Excellent for extended conversations
  • Clear pronunciation and intonation
  • Slightly higher latency (~180-200ms) due to quality

​
When to Use

  • Best audio quality required
  • Extended voice conversations
  • Professional use cases
  • Disk space not a constraint (310 MB)

​
Kokoro Multi-lang v1.1

​
Specifications

  • Provider: Hexgrad
  • Architecture: Kokoro (82M parameters)
  • Size: ~500 MB
  • Speakers: 103 (multilingual)
  • Quality: Excellent
  • Languages: Chinese + English
  • License: Apache 2.0
  • Download: rcli voices

​
Model Files

Stored in ~/Library/RCLI/models/kokoro-multi-lang-v1_1/: 
kokoro-multi-lang-v1_1/
├── model.onnx
├── tokens.txt
├── voices.bin
└── lexicon-us-en.txt

​
Key Features

  • 103 speakers — Largest voice library
  • Chinese and English support
  • Lexicon-based pronunciation (US English)
  • Best for multilingual use cases

​
When to Use

  • Multilingual conversations (Chinese/English)
  • Need many voice options
  • Maximum voice variety
  • Disk space not a constraint (500 MB)

​
Quality Comparison

​
Naturalness & Prosody

VoiceNaturalnessProsodyClarityBest For
Kokoro EnglishExcellentExcellentExcellentExtended conversations
Kokoro Multi-langExcellentExcellentExcellentMultilingual use
KittenTTS NanoGreatGreatGoodVoice variety (8 speakers)
Matcha LJSpeechGreatGoodExcellentClear female voice
Piper AmyGoodGoodGoodWarm female tone
Piper LessacGoodGoodGoodFast, neutral male voice

​
Latency Benchmarks

Measured on Apple M3 Max:
VoiceAvg LatencySynthesis SpeedReal-time Factor
Piper Lessac ⭐150.6 msFast~0.3x
Piper Amy155 msFast~0.3x
KittenTTS Nano165 msFast~0.35x
Matcha LJSpeech170 msFast~0.38x
Kokoro English190 msMedium~0.42x
Kokoro Multi-lang210 msMedium~0.48x
Real-time factor <1.0 means synthesis is faster than playback. All models are fast enough for real-time conversation.

​
Switching Voices

​
Interactive Voice Browser

rcli voices  # Browse, download, and switch TTS voices
Use arrow keys to select a voice, press Enter to download/activate.

​
Check Active Voice

rcli info  # Shows active TTS voice

​
Manual Configuration

Edit config file directly: 
# ~/.Library/RCLI/config
tts_model=kokoro-en      # Voice ID
tts_speaker=5            # Speaker ID (for multi-speaker models)
Changes take effect immediately (no restart required).

​
Architecture Details

​
VITS (Piper)

  • Full name: Variational Inference with adversarial learning for end-to-end TTS
  • Approach: End-to-end neural TTS with GAN-based training
  • Strengths: Fast inference, small model size
  • Phonemizer: eSpeak-NG (built-in)

​
Kokoro

  • Parameters: 82M
  • Approach: Transformer-based TTS with multi-speaker embeddings
  • Strengths: Best naturalness, excellent prosody
  • Phonemizer: eSpeak-NG + lexicon

​
Kitten

  • Architecture: Custom lightweight TTS
  • Approach: Fast synthesis with speaker embeddings
  • Strengths: 8 voices in one model, lightweight

​
Matcha

  • Architecture: Flow-matching for fast TTS
  • Vocoder: HiFi-GAN v2
  • Strengths: Fast synthesis, clear output

​
eSpeak Phonemizer

All TTS models use eSpeak-NG for text-to-phoneme conversion:
  • Size: ~8 MB (eSpeak-NG data)
  • Location: ~/Library/RCLI/models/espeak-ng-data/
  • Languages: 100+ languages supported
  • Shared: All TTS models use the same eSpeak data

​
Benchmarks

Run TTS benchmarks: 
rcli bench --suite tts            # Benchmark active voice
rcli bench --all-tts --suite tts  # Compare all installed voices
Example output: 
=== TTS Benchmark ===
Voice: Piper Lessac
  Average latency: 150.6 ms
  Synthesis speed: 3.2x real-time
  Memory usage: 180 MB

​
Storage Requirements

SetupTotal SizeVoices Installed
Default~68 MBPiper Lessac + eSpeak
+ Kokoro English~378 MBPiper + Kokoro EN
+ Kokoro Multi~878 MBPiper + Both Kokoro
All voices~1.1 GBAll 6 voices + eSpeak
Multiple voices can be installed simultaneously. RCLI selects the active voice from config.

​
Speaker IDs

​
KittenTTS Nano (8 speakers)

Speaker IDGenderDescription
0MaleDeep, calm
1MaleNeutral
2MaleWarm
3MaleEnergetic
4FemaleSoft
5FemaleNeutral
6FemaleWarm
7FemaleClear

​
Kokoro English (11 speakers)

Speaker IDs 0-10 with varied tones and styles. Browse in TUI for previews.

​
Kokoro Multi-lang (103 speakers)

Speaker IDs 0-102 covering Chinese and English voices. Use TUI to explore.

​
Next Steps

Voice Browser

Browse and download TTS voices

Benchmarks

Measure TTS performance on your Mac

Model Browser

Manage all models (LLM, STT, TTS)

Switch Models

Hot-swap voices without restart

Build docs developers (and LLMs) love

Get started for freeTalk to us