Instructions to use tokenaii/Replica-Preview-0.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Chatterbox
How to use tokenaii/Replica-Preview-0.5B with Chatterbox:
# pip install chatterbox-tts import torchaudio as ta from chatterbox.tts import ChatterboxTTS model = ChatterboxTTS.from_pretrained(device="cuda") text = "Ezreal and Jinx teamed up with Ahri, Yasuo, and Teemo to take down the enemy's Nexus in an epic late-game pentakill." wav = model.generate(text) ta.save("test-1.wav", wav, model.sr) # If you want to synthesize with a different voice, specify the audio prompt AUDIO_PROMPT_PATH="YOUR_FILE.wav" wav = model.generate(text, audio_prompt_path=AUDIO_PROMPT_PATH) ta.save("test-2.wav", wav, model.sr) - Notebooks
- Google Colab
- Kaggle
Replica Preview 0.5B
Replica Preview is a text-to-speech (TTS) model family from TokenAI. Replica Preview 0.5B is a multilingual, expressive TTS model β an autoregressive speech-token language model with a flow-matching vocoder β with a multilingual focus. It supports zero-shot voice cloning from a short reference recording and expressive speech control through sampling and expression parameters.
Replica Preview 0.5B is part of the Replica Preview model family by TokenAI. Learn more about the family: https://tokenai.llc/projects/replica-preview β official model page: https://tokenai.llc/projects/models/replica-preview-0.5b
Model Details
| Property | Value |
|---|---|
| Model type | Text-to-Speech (TTS), autoregressive speech-token LM + flow-matching vocoder |
| Architecture | T3 multilingual (~520M) speech-token LM + S3Gen vocoder |
| Speech-token vocabulary | 8,194 tokens |
| Speaker embedding | 256 dimensions |
| Maximum text context | 2,048 text tokens |
| Sampling controls | temperature, top_p, min_p, repetition_penalty, cfg_weight, exaggeration |
| Expression defaults | exaggeration = 0.5, cfg_weight = 0.5 |
| Dramatic expression | exaggeration β 0.7 (or higher), cfg_weight β 0.3 |
| Sample rate | 24 kHz (mono output, verified in benchmark) |
| Languages | 17: Arabic, German, English, Spanish, French, Hindi, Italian, Japanese, Korean, Malay, Dutch, Norwegian, Polish, Portuguese, Russian, Turkish, Chinese |
| Reference conditioning | ~6β10 seconds of clean, clear, single-speaker audio |
| License | MIT |
Capabilities
- Zero-shot voice cloning: clone any speaker from a short reference recording (typically 6β10 s of clean, single-speaker audio); no fixed speaker-ID list
- Expressive speech synthesis: control delivery via
exaggerationandcfg_weight(defaults 0.5 / 0.5; dramatic delivery β 0.7 / 0.3) - Multilingual generation: 17 languages with language-ID conditioning
- Long-context text: up to 2,048 text tokens
Performance
English inference and output-integrity benchmark (2026-11-13, 8 fixed English prompts, seeded, single request, NVIDIA RTX PRO 6000 Blackwell Server Edition):
| Metric | Result |
|---|---|
| English prompts | 8 |
| Total generated audio | 57.96 s |
| Total generation time | 26.65 s |
| Mean generation time | 3.331 s (median 3.205 s, P95 5.028 s) |
| Mean real-time factor (RTF) | 0.463 |
| Approximate throughput | ~2.22x real-time |
| Model load time | 9.280 s |
| Output format | 24 kHz mono PCM16 WAV |
| Peak allocated GPU memory | 3.809 GB (3.548 GiB) |
| Peak reserved GPU memory | 4.100 GB |
| Non-finite samples (NaN/Inf) | 0 |
| Clipped samples | 0 |
Output-integrity checks passed 8/8: WAV generation, sample rate, mono layout, finite samples, no hard clipping, no empty files, no runtime exceptions. A warm-up generation was run before the timed samples.
Teacher-forced cross-entropy (not MOS, WER, CER, or human naturalness scores):
| Evaluation set | Text CE | Speech CE |
|---|---|---|
| English | 1.907 | 3.835 |
| Combined Arabic | 1.744 | 4.278 |
| Egyptian-only | 1.717 | 4.051 |
These results establish that the model loads, synthesizes the tested English prompts, and produces valid 24 kHz mono audio at the measured speed without numerical failure or clipping. They do not establish a human quality score, transcript fidelity, or production capacity under concurrency.
Latency
| Scenario | Value |
|---|---|
| Mean generation time | 3.331 s |
| Median / P95 generation time | 3.205 s / 5.028 s |
| Mean RTF | 0.463 (~2.22x real-time) |
| Model load time | 9.280 s |
Memory
| Scenario | VRAM |
|---|---|
| Inference (single request, bf16, CUDA) | 3.809 GB allocated (4.100 GB reserved) |
Measured on NVIDIA RTX PRO 6000 Blackwell Server Edition (CUDA only; CPU, batch, and low-VRAM configurations not measured).
Usage
Replica Preview 0.5B generates expressive speech with optional zero-shot cloning from a ~6β10 s reference recording. Sampling is controlled via temperature, top_p, min_p and repetition_penalty; delivery is shaped with cfg_weight and exaggeration (defaults 0.5 / 0.5).
exaggeration = 0.5 # raise toward 0.7+ for dramatic delivery
cfg_weight = 0.5 # lower toward 0.3 for dramatic delivery
Note: the model weights for this repository are being packaged for release. Until they are published here, see the official model page for access: https://tokenai.llc/projects/models/replica-preview-0.5b
Audio Samples
Audio samples are not hosted or displayed in this repository. To hear the official audio samples β and to see the full technical details and all features of this model β visit the official Replica Preview 0.5B page on TokenAI's website: https://tokenai.llc/projects/models/replica-preview-0.5b
Limitations
- Benchmark coverage is English-only (8 fixed prompts); Arabic and cross-lingual inference are not measured in this report
- No human MOS / naturalness scoring yet; native listening remains pending
- ASR WER/CER not measured; teacher-forced CE is not an intelligibility score
- Voice-clone similarity is not certified in this run (voice-encoder proxy unavailable); no speaker-verification claim is made
- Long-form continuity (>10 s), concurrency (1/2/4/8 requests), batch inference, CPU inference, and low-VRAM behavior were not measured
- English teacher-forced CE is effectively flat (1.905 β 1.907); do not read it as an English improvement
- Cloning quality depends on reference cleanliness, single-speaker consistency, language match, and pronunciation clarity
- Requires the T3 + S3Gen runtime components
About TokenAI
TokenAI is a non-profit AI startup founded on May 10, 2025 by Assem Sabry. TokenAI builds advanced AI models, research, datasets, and developer tools. We focus on language, multimodal, speech, cybersecurity, and Arabic-native intelligence. Our mission is to make advanced intelligence accessible and useful for everyone.
Website: https://assem.one
Contact TokenAI
For inquiries, partnerships, or support, contact TokenAI at info@tokenai.llc.
Citation
@misc{replica_preview_05b_2026,
title={Replica Preview 0.5B: Multilingual TTS with Zero-Shot Voice Cloning},
author={TokenAI},
year={2026},
url={https://huggingface.co/tokenaii/Replica-Preview-0.5B}
}
License
MIT. See LICENSE for details.
- Downloads last month
- -