​

Other TTS Models

​

Overview

​

KittenTTS

​

Description

​

Characteristics

• Streaming: ✅ Yes
• Quality: ⭐⭐⭐ Good
• Speed: ⭐⭐⭐⭐⭐ Very Fast
• Memory: ⭐⭐⭐⭐⭐ Very Low
• Size: Very Small (typically 10-30 MB)
• Multi-Speaker: ✅ Yes

​

Configuration

import { createTTS } from 'react-native-sherpa-onnx/tts';

const tts = await createTTS({
  modelPath: {
    type: 'asset',
    path: 'models/kitten-tts-en'
  },
  modelType: 'kitten', // or 'auto'
  numThreads: 2,
});

const audio = await tts.generateSpeech('Hello from KittenTTS!');
console.log('Generated:', audio.samples.length, 'samples');

await tts.destroy();

​

Streaming Example

const sampleRate = await tts.getSampleRate();
await tts.startPcmPlayer(sampleRate, 1);

await tts.generateSpeechStream('Fast streaming speech', { sid: 0, speed: 1.0 }, {
  onChunk: async (chunk) => {
    await tts.writePcmChunk(chunk.samples);
  },
  onEnd: async () => {
    await tts.stopPcmPlayer();
  },
});

​

Download

​

Model Detection

• Folder name should contain kitten (not kokoro)
• Files: model.onnx, tokens.txt

​

When to Use

​

Advantages

1. Very Fast: Fastest TTS model available
2. Very Small: Minimal storage footprint
3. Low Memory: Runs on constrained devices
4. Streaming: Low-latency incremental generation
5. Multi-Speaker: Multiple voices in one model

​

Limitations

1. Quality: Good but not as natural as VITS or Matcha
2. Limited Languages: Fewer language options
3. No Voice Cloning: Fixed voice set only

​

Zipvoice

​

Description

​

Characteristics

• Streaming: ❌ No (batch only for voice cloning)
• Quality: ⭐⭐⭐⭐⭐ Excellent
• Speed: ⭐⭐⭐ Medium
• Memory: ⭐⭐ High (requires significant RAM)
• Size: Large (~605 MB for full model)
• Voice Cloning: ✅ Yes

​

Architecture

• Encoder – Encodes reference audio
• Decoder (flow-matching) – Generates mel-spectrogram
• Vocoder (e.g. vocos_24khz.onnx) – Converts to waveform

​

Configuration

import { createTTS } from 'react-native-sherpa-onnx/tts';

const tts = await createTTS({
  modelPath: {
    type: 'asset',
    path: 'models/sherpa-onnx-zipvoice-zh-en-emilia'
  },
  modelType: 'zipvoice',
  numThreads: 4,
});

// Voice cloning with reference audio
const audio = await tts.generateSpeech('Target text to speak in reference voice', {
  referenceAudio: {
    samples: referenceSamples,  // float[] from reference audio
    sampleRate: 22050,
  },
  referenceText: 'Transcript of the reference audio',
  numSteps: 20,  // Flow-matching steps (higher = better quality)
  speed: 1.0,
});

await tts.destroy();

​

Memory Requirements

​

Download

​

Model Detection

• Encoder + decoder + vocoder files
• Optional folder name pattern (containing zipvoice)
• Files: encoder, decoder, vocos_*.onnx (vocoder), tokens.txt, lexicon.txt, espeak-ng-data

​

When to Use

​

Advantages

1. Zero-Shot Voice Cloning: Clone any voice from short audio
2. Excellent Quality: Very natural-sounding output
3. Flexible: Works with various reference voices
4. Multilingual: Supports Chinese and English

​

Limitations

1. High Memory: Full model needs 8+ GB device RAM
2. No Streaming: Voice cloning only supports batch generation
3. Large Size: ~605 MB (use int8 distill variant for smaller size)
4. Slower: Flow-matching is computationally intensive
5. Requires Vocoder: Distill-only models (no vocoder) will fail

​

Reference Audio Requirements

• Format: Mono, float PCM samples in [-1, 1]
• Sample Rate: Typically 22050 Hz or 24000 Hz
• Duration: 3-10 seconds recommended
• Quality: Clear speech, minimal background noise
• Transcript: Must provide accurate transcript of reference audio

​

Pocket

​

Description

​

Characteristics

• Streaming: ✅ Yes (including with reference audio for Kotlin-engine models)
• Quality: ⭐⭐⭐⭐ High
• Speed: ⭐⭐⭐⭐ Fast
• Memory: ⭐⭐⭐ Moderate
• Size: Medium
• Voice Cloning: ✅ Yes

​

Configuration

import { createTTS } from 'react-native-sherpa-onnx/tts';

const tts = await createTTS({
  modelPath: {
    type: 'asset',
    path: 'models/pocket-tts'
  },
  modelType: 'pocket', // or 'auto'
  numThreads: 2,
});

// Standard generation
const audio1 = await tts.generateSpeech('Hello from Pocket TTS');

// Voice cloning with reference audio
const audio2 = await tts.generateSpeech('Target text', {
  referenceAudio: { samples: refSamples, sampleRate: 22050 },
  referenceText: 'Reference transcript',
  numSteps: 20,
  extra: {
    temperature: '0.7',
    chunk_size: '15',
  },
});

await tts.destroy();

​

Streaming with Voice Cloning

await tts.generateSpeechStream('Target text', {
  referenceAudio: { samples: refSamples, sampleRate: 22050 },
  referenceText: 'Reference transcript',
  numSteps: 20,
}, {
  onChunk: async (chunk) => {
    await tts.writePcmChunk(chunk.samples);
  },
});

​

Extra Options

const audio = await tts.generateSpeech('Text', {
  sid: 0,
  speed: 1.0,
  numSteps: 20,
  extra: {
    temperature: '0.7',    // Sampling temperature
    chunk_size: '15',      // Processing chunk size
    // Add other model-specific options as needed
  },
});

​

Download

​

Model Detection

• Files: lm_flow, lm_main, text_conditioner, vocab/token_scores
• No folder name pattern required

​

When to Use

​

Advantages

1. Streaming + Voice Cloning: Supports both simultaneously
2. Flow-Matching: Modern architecture for quality
3. Fast: Good performance with streaming
4. Flexible: Extra options for fine-tuning
5. Good Quality: Natural-sounding speech

​

Limitations

1. Newer: Less battle-tested than VITS or Zipvoice
2. Documentation: Fewer examples and resources
3. Model Availability: Fewer pretrained models

​

Comparison Table

​

Choosing Between Models

​

For Voice Cloning

• Zipvoice – Best quality, batch generation only, high memory
• Pocket – Streaming support, good quality, moderate memory

​

For Speed

• KittenTTS – Fastest, lightweight
• Pocket – Fast with streaming

​

For Low-End Devices

• KittenTTS – Minimal resources
• Zipvoice int8 distill – If voice cloning is needed

​

For High Quality

• Zipvoice – Excellent voice cloning quality
• Pocket – Good quality with more flexibility

​

Next Steps