ResembleAI/chatterbox-turbo-ONNX

Chatterbox TTS

Made with ❤️ by

Chatterbox is a family of three state-of-the-art, open-source text-to-speech models by Resemble AI.

We are excited to introduce Chatterbox-Turbo, our most efficient model yet. Built on a streamlined 350M parameter architecture, Turbo delivers high-quality speech with less compute and VRAM than our previous models. We have also distilled the speech-token-to-mel decoder, previously a bottleneck, reducing generation from 10 steps to just one, while retaining high-fidelity audio output.

Paralinguistic tags are now native to the Turbo model, allowing you to use [cough], [laugh], [chuckle], and more to add distinct realism. While Turbo was built primarily for low-latency voice agents, it excels at narration and creative workflows.

If you like the model but need to scale or tune it for higher accuracy, check out our competitively priced TTS service (link). It delivers reliable performance with ultra-low latency of sub 200ms—ideal for production use in agents, applications, or interactive media.

⚡ Model Zoo

Choose the right model for your application.

Model	Size	Languages	Key Features	Best For	🤗	Examples
Chatterbox-Turbo	350M	English	Paralinguistic Tags ([laugh]), Lower Compute and VRAM	Zero-shot voice agents, Production	Demo	Listen
Chatterbox-Multilingual (Language list)	500M	23+	Zero-shot cloning, Multiple Languages	Global applications, Localization	Demo	Listen
Chatterbox (Tips and Tricks)	500M	English	CFG & Exaggeration tuning	General zero-shot TTS with creative controls	Demo	Listen

Usage

Chatterbox-Turbo

import onnxruntime
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
import numpy as np
from tqdm import trange
import librosa
import soundfile as sf

MODEL_ID = "ResembleAI/chatterbox-turbo-ONNX"
SAMPLE_RATE = 24000
START_SPEECH_TOKEN = 6561
STOP_SPEECH_TOKEN = 6562
SILENCE_TOKEN = 4299
NUM_KV_HEADS = 16
HEAD_DIM = 64

class RepetitionPenaltyLogitsProcessor:
    def __init__(self, penalty: float):
        if not isinstance(penalty, float) or not (penalty > 0):
            raise ValueError(f"`penalty` must be a strictly positive float, but is {penalty}")
        self.penalty = penalty

    def __call__(self, input_ids: np.ndarray, scores: np.ndarray) -> np.ndarray:
        score = np.take_along_axis(scores, input_ids, axis=1)
        score = np.where(score < 0, score * self.penalty, score / self.penalty)
        scores_processed = scores.copy()
        np.put_along_axis(scores_processed, input_ids, score, axis=1)
        return scores_processed

def download_model(name: str, dtype: str = "fp32") -> str:
    filename = f"{name}{'' if dtype == 'fp32' else '_quantized' if dtype == 'q8' else f'_{dtype}'}.onnx"
    graph = hf_hub_download(MODEL_ID, subfolder="onnx", filename=filename)      # Download graph
    hf_hub_download(MODEL_ID, subfolder="onnx", filename=f"{filename}_data")    # Download weights
    return graph

# Download models
## dtype options: fp32, fp16, q8, q4, q4f16
conditional_decoder_path = download_model("conditional_decoder", dtype="fp32")
speech_encoder_path = download_model("speech_encoder", dtype="fp32")
embed_tokens_path = download_model("embed_tokens", dtype="fp32")
language_model_path = download_model("language_model", dtype="fp32")

# Create ONNX sessions
speech_encoder_session = onnxruntime.InferenceSession(speech_encoder_path)
embed_tokens_session = onnxruntime.InferenceSession(embed_tokens_path)
language_model_session = onnxruntime.InferenceSession(language_model_path)
cond_decoder_session = onnxruntime.InferenceSession(conditional_decoder_path)

# Generation parameters
text = "Oh, that's hilarious! [chuckle] Um anyway, how are you doing today?"
target_voice_path = "path/to/voice.wav"
output_file_name = "output.wav"
max_new_tokens = 1024
repetition_penalty = 1.2
apply_watermark = False

# Prepare audio input
audio_values, _ = librosa.load(target_voice_path, sr=SAMPLE_RATE)
audio_values = audio_values[np.newaxis, :].astype(np.float32)

# Prepare text input
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
input_ids = tokenizer(text, return_tensors="np")["input_ids"].astype(np.int64)

# Generation loop
repetition_penalty_processor = RepetitionPenaltyLogitsProcessor(penalty=repetition_penalty)
generate_tokens = np.array([[START_SPEECH_TOKEN]], dtype=np.int64)
for i in trange(max_new_tokens, desc="Sampling", dynamic_ncols=True):
    inputs_embeds = embed_tokens_session.run(None, {"input_ids": input_ids})[0]

    if i == 0:
        ort_speech_encoder_input = {"audio_values": audio_values}
        cond_emb, prompt_token, speaker_embeddings, speaker_features = speech_encoder_session.run(None, ort_speech_encoder_input)
        inputs_embeds = np.concatenate((cond_emb, inputs_embeds), axis=1)

        # Initialize cache and LLM inputs
        batch_size, seq_len, _ = inputs_embeds.shape
        past_key_values = {
            i.name: np.zeros([batch_size, NUM_KV_HEADS, 0, HEAD_DIM], dtype=np.float16 if i.type == 'tensor(float16)' else np.float32)
            for i in language_model_session.get_inputs()
            if "past_key_values" in i.name
        }
        attention_mask = np.ones((batch_size, seq_len), dtype=np.int64)
        position_ids = np.arange(seq_len, dtype=np.int64).reshape(1, -1).repeat(batch_size, axis=0)

    logits, *present_key_values = language_model_session.run(None, dict(
        inputs_embeds=inputs_embeds,
        attention_mask=attention_mask,
        position_ids=position_ids,
        **past_key_values,
    ))

    logits = logits[:, -1, :]
    next_token_logits = repetition_penalty_processor(generate_tokens, logits)

    input_ids = np.argmax(next_token_logits, axis=-1, keepdims=True).astype(np.int64)
    generate_tokens = np.concatenate((generate_tokens, input_ids), axis=-1)
    if (input_ids.flatten() == STOP_SPEECH_TOKEN).all():
        break

    # Update values for next generation loop
    attention_mask = np.concatenate([attention_mask, np.ones((batch_size, 1), dtype=np.int64)], axis=1)
    position_ids = position_ids[:, -1:] + 1
    for j, key in enumerate(past_key_values):
        past_key_values[key] = present_key_values[j]

# Decode audio
speech_tokens = generate_tokens[:, 1:-1]
silence_tokens = np.full((speech_tokens.shape[0], 3), SILENCE_TOKEN, dtype=np.int64) # Add silence at the end
speech_tokens = np.concatenate([prompt_token, speech_tokens, silence_tokens], axis=1)

wav = cond_decoder_session.run(None, dict(
    speech_tokens=speech_tokens,
    speaker_embeddings=speaker_embeddings,
    speaker_features=speaker_features,
))[0].squeeze(axis=0)

# Optional: Apply watermark
if apply_watermark:
    import perth
    watermarker = perth.PerthImplicitWatermarker()
    wav = watermarker.apply_watermark(wav, sample_rate=SAMPLE_RATE)

sf.write(output_file_name, wav, SAMPLE_RATE)

Built-in PerTh Watermarking for Responsible AI

Every audio file generated by Chatterbox includes Resemble AI's Perth (Perceptual Threshold) Watermarker - imperceptible neural watermarks that survive MP3 compression, audio editing, and common manipulations while maintaining nearly 100% detection accuracy.

Watermark extraction

You can look for the watermark using the following script.

import perth
import librosa

AUDIO_PATH = "YOUR_FILE.wav"

# Load the watermarked audio
watermarked_audio, sr = librosa.load(AUDIO_PATH, sr=None)

# Initialize watermarker (same as used for embedding)
watermarker = perth.PerthImplicitWatermarker()

# Extract watermark
watermark = watermarker.get_watermark(watermarked_audio, sample_rate=sr)
print(f"Extracted watermark: {watermark}")
# Output: 0.0 (no watermark) or 1.0 (watermarked)

Official Discord

👋 Join us on Discord and let's build something awesome together!

Acknowledgements

• Cosyvoice
• Real-Time-Voice-Cloning
• HiFT-GAN
• Llama 3
• S3Tokenizer

Citation

If you find this model useful, please consider citing.

@misc{chatterboxtts2025,
  author       = {{Resemble AI}},
  title        = {{Chatterbox-TTS}},
  year         = {2025},
  howpublished = {\url{https://github.com/resemble-ai/chatterbox}},
  note         = {GitHub repository}
}

Disclaimer

Don't use this model to do bad things. Prompts are sourced from freely available data on the internet.