chatterbox-ONNX

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.

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