Kokoro-7M-ONNX

Kokoro-7M ONNX Models & Voice Pack

Ultra-lightweight (7.48M parameters) Text-to-Speech bundle for Kokoro-7M-Distill (oddadmix/Kokoro-7M-Distill), containing both Full FP32 and INT8 Quantized ONNX models, fully optimized for CPU execution with ONNX Runtime and paired with the official conditioned voice af_msa.onnx.

• Sample Rate: 24 kHz
• Conditioned Voice: af_msa.onnx
• CPU Speed: 17.6x–18.4x faster than real-time (~330 ms for 6.1s speech)
• Zero Weights at Inference: 100% pure ONNX Runtime execution.

---

Files in this Folder

[!IMPORTANT] Why af_msa? The student model was distilled and conditioned specifically on the af_msa voice pack. Distillation benchmarks show using af_heart severely degrades naturalness and clarity (UTMOS drops from 4.14 to 3.57). Always pair Kokoro-7M with af_msa.

---

Quickstart: How to Test

1. Requirements

Install minimal inference dependencies:

pip install onnxruntime soundfile numpy kokoro

2. Run Test

Test Full Precision Model (Default):

python test.py

Test INT8 Quantized Model:

python test.py --int8

Test Custom Text:

python test.py --text "Hello world! Testing Kokoro seven million on CPU."
python test.py --int8 --text "Testing INT8 quantized speech synthesis."

Generated audio is saved to out.wav (24 kHz, 16-bit PCM).

---

Performance (CPU Benchmarks)

Benchmarked on an Intel Core CPU with standard 4-thread budget:

---

Python Usage (100% Pure ONNX Runtime)

import numpy as np
import onnxruntime as ort
import soundfile as sf
from kokoro import KPipeline

# 1. Choose model: "kokoro_7m.onnx" (full) or "kokoro_7m_int8.onnx" (int8)
model_path = "kokoro_7m.onnx"  # or "kokoro_7m_int8.onnx"

sess_opts = ort.SessionOptions()
sess_opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

model_sess = ort.InferenceSession(model_path, sess_opts, providers=["CPUExecutionProvider"])
voice_sess = ort.InferenceSession("af_msa.onnx", sess_opts, providers=["CPUExecutionProvider"])

# 2. Phonemize text
pipeline = KPipeline(lang_code="a", repo_id="oddadmix/Kokoro-7M-Distill", model=False)
text = "Hello! Kokoro-7M is running with an ONNX voice on CPU."
_, tokens = pipeline.g2p(text)

# 114-symbol vocabulary mapping
VOCAB = {
    ';': 1, ':': 2, ',': 3, '.': 4, '!': 5, '?': 6, '—': 9, '…': 10, '"': 11,
    '(': 12, ')': 13, '“': 14, '”': 15, ' ': 16, '̃': 17, 'ʣ': 18, 'ʥ': 19,
    'ʦ': 20, 'ʨ': 21, 'ᵝ': 22, 'ꭧ': 23, 'A': 24, 'I': 25, 'O': 31, 'Q': 33,
    'S': 35, 'T': 36, 'W': 39, 'Y': 41, 'ᵊ': 42, 'a': 43, 'b': 44, 'c': 45,
    'd': 46, 'e': 47, 'f': 48, 'h': 50, 'i': 51, 'j': 52, 'k': 53, 'l': 54,
    'm': 55, 'n': 56, 'o': 57, 'p': 58, 'q': 59, 'r': 60, 's': 61, 't': 62,
    'u': 63, 'v': 64, 'w': 65, 'x': 66, 'y': 67, 'z': 68, 'ɑ': 69, 'ɐ': 70,
    'ɒ': 71, 'æ': 72, 'β': 75, 'ɔ': 76, 'ɕ': 77, 'ç': 78, 'ɖ': 80, 'ð': 81,
    'ʤ': 82, 'ə': 83, 'ɚ': 85, 'ɛ': 86, 'ɜ': 87, 'ɟ': 90, 'ɡ': 92, 'ɥ': 99,
    'ɨ': 101, 'ɪ': 102, 'ʝ': 103, 'ɯ': 110, 'ɰ': 111, 'ŋ': 112, 'ɳ': 113,
    'ɲ': 114, 'ɴ': 115, 'ø': 116, 'ɸ': 118, 'θ': 119, 'œ': 120, 'ɹ': 123,
    'ɾ': 125, 'ɻ': 126, 'ʁ': 128, 'ɽ': 129, 'ʂ': 130, 'ʃ': 131, 'ʈ': 132,
    'ʧ': 133, 'ʊ': 135, 'ʋ': 136, 'ʌ': 138, 'ɣ': 139, 'ɤ': 140, 'χ': 142,
    'ʎ': 143, 'ʒ': 147, 'ʔ': 148, 'ˈ': 156, 'ˌ': 157, 'ː': 158, 'ʰ': 162,
    'ʲ': 164, '↓': 169, '→': 171, '↗': 172, '↘': 173, 'ᵻ': 177
}

# 3. Synthesize
chunks = []
silence = np.zeros(int(24000 * 0.15), dtype=np.float32)

for _, ps, _ in pipeline.en_tokenize(tokens):
    if not ps: continue
    input_ids = np.array([[0] + [VOCAB[p] for p in ps if p in VOCAB] + [0]], dtype=np.int64)

    # Query style vector from voice ONNX model
    idx = np.array([min(input_ids.shape[1] - 1, 509)], dtype=np.int64)
    style_vec = voice_sess.run(None, {"index": idx})[0]

    # Generate audio chunk from model ONNX
    chunk = model_sess.run(None, {
        "input_ids": input_ids,
        "style": style_vec.astype(np.float32),
        "speed": np.array([1.0], dtype=np.float32)
    })[0].squeeze()

    chunks.append(chunk)
    chunks.append(silence)

full_audio = np.concatenate(chunks[:-1]) if len(chunks) > 1 else chunks[0]
sf.write("out.wav", full_audio, 24000)
print(f"Generated {len(full_audio)/24000:.2f}s of audio saved to out.wav!")