Replica Preview 0.6B

Replica Preview 0.6B

Replica Preview is the first text-to-speech (TTS) model from TokenAI and the beginning of the Replica model family. It is a non-autoregressive, diffusion-based multilingual TTS system built on top of OmniVoice, with a focus on natural, expressive speech and non-verbal sounds (breaths, sighs, laughter, fillers).

Replica Preview 0.6B is part of the Replica Preview model family by TokenAI. Learn more about the family: https://tokenai.llc/projects/replica-preview

Model Details

Property Value
Model type Text-to-Speech (TTS), non-autoregressive diffusion
Base model OmniVoice (k2-fsa/OmniVoice)
LLM Qwen/Qwen3-0.6B-Base
Audio codec HiggsAudioV2Tokenizer: 9 codebooks x 1024, hop 960, 16 kHz in / 24 kHz out
Audio heads 8 heads with weights [8, 8, 6, 6, 4, 4, 2, 2]
Parameters 612.6M (LLM 595.8M + audio embeddings 8.4M + audio heads 8.4M)
Model size 2.45 GB (safetensors, float32) + 0.77 GB codec = 3.22 GB total
Sample rate 24 kHz output
Languages 646 (inherited from OmniVoice); optimized for Egyptian Arabic (arz), English, French, German, Spanish
License Apache 2.0

Capabilities

  • Zero-shot voice cloning: clone any speaker from a short reference audio (unlimited speakers)
  • Voice Design: gender x age x pitch x style x accent
  • Auto voice: automatic voice selection
  • Non-verbal sounds: [inhale], [exhale], [sigh], [laughter], [laugh], and fillers (ู‡ู…ู…ู…, ุขู‡, ah, oh, Hmmm, phew)
  • Number normalization: Arabic digits are converted to Egyptian words (e.g., 5000 -> ุฎู…ุณุฉ ุขู„ุงู)
  • Streaming: sentence-chunked streaming with RTF 0.153 (~6.5x realtime)
  • Multilingual: 646 languages inherited, optimized for Egyptian Arabic + code-switching

Performance

Evaluated via round-trip CER (generate speech, transcribe with Whisper large-v3, compare to input):

Metric Value
Egyptian Arabic CER 0.066 - 0.086
Conversational CER 0.0937
English CER 0.041
Code-switch CER 0.2488 - 0.270 (weakest area)
Numbers CER 0.00 - 0.26
Fail rate (CER > 0.2) 0.03
RTF (streaming) 0.153 (~6.5x realtime)
TTFA (time-to-first-audio) ~0.65s
DNSMOS ovrl_mos 3.35 (Arabic) / 3.47 (English)
MFCC speaker similarity 0.9985 (Arabic) / 0.9972 (English)
Voice Design fidelity 0.999

Latency

Scenario Value
First-chunk latency (streaming) ~0.65s
Full sentence (~30s audio) ~2-4s wall time
RTF (streaming, sentence-chunked) 0.153
RTF (batched, num_step=8) ~0.01 (very fast)

Memory

Scenario VRAM
Inference (bf16) ~2.5-3 GB

Usage

Voice cloning (recommended for production)

from omnivoice import OmniVoice
import soundfile as sf

model = OmniVoice.from_pretrained("tokenaii/Replica-Preview", device_map="cuda:0", dtype="bfloat16")
prompt = model.create_voice_clone_prompt(ref_audio="reference.wav", ref_text="transcript of reference")
audio = model.generate(text="ู…ุฑุญุจุงุŒ ุฃู†ุง ุฑูุจู„ูŠูƒุง.", language_id="arz", voice_clone_prompt=prompt, num_step=32)
sf.write("output.wav", audio[0], 24000)

Voice Design

audio = model.generate(text="Hello, I am Replica.", language_id="en",
                      instruct="female, young adult, american accent", num_step=32)

Non-verbal sounds

Include markers in the text: [inhale], [exhale], [sigh], [laughter], [laugh], and fillers like ู‡ู…ู…ู…, ุขู‡, Hmmm, phew.

Number normalization

Arabic digits are automatically converted to Egyptian word forms. Example: "5000" -> "ุฎู…ุณุฉ ุขู„ุงู".

Audio Samples

Audio samples are not hosted or displayed in this repository. To hear the official audio samples โ€” and to see the full technical details and all features of this model โ€” visit the official Replica Preview 0.6B page on TokenAI's website: https://tokenai.llc/projects/models/replica-preview-0.6b

Limitations

  • Code-switching (Arabic-English mixing) is the weakest area (CER 0.25)
  • Number pronunciation occasionally varies ("ุฎู…ุณ ุชู„ุงู" vs "ุฎู…ุณุฉ ุขู„ุงู")
  • Requires the Higgs Audio V2 tokenizer (included in audio_tokenizer/)

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_2026,
  title={Replica Preview 0.6B: Egyptian Arabic TTS with Code-Switching},
  author={TokenAI},
  year={2026},
  url={https://huggingface.co/tokenaii/Replica-Preview-0.6B}
}

License

Apache 2.0. See LICENSE for details.

Downloads last month
22
Safetensors
Model size
0.6B params
Tensor type
I64
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tokenaii/Replica-Preview-0.6B

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Finetuned
(86)
this model

Collection including tokenaii/Replica-Preview-0.6B