spark-tts

name

spark-tts

description

Generate speech from text using iFlytek's Spark TTS model locally on Apple Silicon via mlx-audio. Supports Chinese and English with controllable gender, pitch, and speed. Also supports voice cloning from a 3-second reference audio clip. Use when: user asks to "generate Chinese speech", "中文语音合成", "TTS in Chinese", "spark tts", "clone my voice", "read this in Chinese", or needs Chinese text-to-speech locally. Preferred over Voxtral for Chinese/CJK content. Lightweight 0.5B model (~1GB).

Spark TTS

Local Chinese+English text-to-speech on Apple Silicon using iFlytek's Spark-TTS-0.5B via MLX.

Requirements

• Apple Silicon Mac (M1+)
• Python 3.10+
• ~1GB RAM (bf16)

Setup

pip install -U mlx-audio

Quick Usage

from mlx_audio.tts.utils import load
import numpy as np
import soundfile as sf

model = load("mlx-community/Spark-TTS-0.5B-bf16")

for result in model.generate(text="你好，世界！", gender="female"):
    audio = np.array(result.audio)
    sf.write("output.wav", audio, result.sample_rate)  # 16000 Hz

Always use result.sample_rate — Spark outputs at 16000 Hz (not 24000). Hardcoding the wrong rate causes audio to sound fast-forwarded.

Parameters

model.generate(
    text="Hello",           # required
    gender="female",        # "male" or "female"
    pitch=1.0,              # 0.0=very_low, 0.5=low, 1.0=moderate, 1.5=high, 2.0=very_high
    speed=1.0,              # 0.0=very_low, 0.5=low, 1.0=moderate, 1.5=high, 2.0=very_high
    temperature=0.8,        # generation temperature
    ref_audio="ref.wav",    # optional: reference audio for voice cloning
    ref_text="transcript",  # optional: transcript of reference audio
)

Gender

• "male" — male voice
• "female" — female voice

Pitch & Speed Levels

Value	Level
0.0	very_low
0.5	low
1.0	moderate (default)
1.5	high
2.0	very_high

Voice Cloning

Clone any voice with ~3 seconds of reference audio:

for result in model.generate(
    text="你好，我是克隆的声音。",
    ref_audio="reference.wav",
    ref_text="这是参考音频的文字内容。",
):
    sf.write("cloned.wav", np.array(result.audio), result.sample_rate)

Script Usage

python scripts/spark_generate.py --text "你好世界" --gender female --output output.wav
python scripts/spark_generate.py --text "Hello" --gender male --pitch 1.5 --speed 1.5 --output fast_high.ogg --format ogg
python scripts/spark_generate.py --text "克隆声音" --ref-audio voice.wav --ref-text "参考文字" --output cloned.wav

Options:

• --text — Text to synthesize (required)
• --gender — male or female (default: female)
• --pitch — Pitch level: 0.0/0.5/1.0/1.5/2.0 (default: 1.0)
• --speed — Speed level: 0.0/0.5/1.0/1.5/2.0 (default: 1.0)
• --ref-audio — Reference audio for voice cloning
• --ref-text — Transcript of reference audio
• --model — Model variant (default: mlx-community/Spark-TTS-0.5B-bf16)
• --output — Output file path (default: output.wav)
• --format — Output format: wav, ogg, mp3 (default: wav)

First Run

The first run downloads the model from HuggingFace (~1GB). The script auto-detects this and prints a notice. Subsequent runs load from cache in ~2 seconds.

Performance (M5 MacBook Pro 32GB)

• Load time: ~2s
• Generation RTF: ~0.63 (10s audio ≈ 6.3s generation)
• Sample rate: 16000 Hz
• Languages: Chinese, English (bilingual)
• Model size: ~1GB

Language Comparison

Feature	Spark TTS	Voxtral TTS
Chinese	✅	❌
English	✅	✅ (better quality)
Other languages	❌	✅ (9 languages)
Voice cloning	✅	❌
Pitch/speed control	✅	❌
Model size	~1GB	~2.5GB
Voices	2 (M/F) × 5 levels	20 presets

Use Spark for Chinese. Use Voxtral for other languages.

Telegram Voice Message

Convert to OGG Opus for Telegram:

ffmpeg -y -i output.wav -c:a libopus -b:a 96k output.ogg

Send with asVoice: true.