cstr/vibevoice-1.5b-GGUF

VibeVoice-1.5B GGUF

GGUF conversion of microsoft/VibeVoice-1.5B for use with CrispASR.

This is the base model (not the streaming variant). It supports voice cloning from audio samples and multi-speaker synthesis.

Model variants

File	Quant	Size	Notes
vibevoice-1.5b-tts-f16.gguf	F16	5.1 GB	Full precision
vibevoice-1.5b-tts-q8_0.gguf	Q8_0	3.1 GB	Near-lossless
vibevoice-1.5b-tts-q4_k.gguf	Q4_K	1.9 GB	Smallest, perfect ASR round-trip

The Q8_0/Q4_K files keep the diffusion prediction head, connectors and EOS classifier at full precision (only the LM backbone is quantized). The head runs under classifier-free guidance, so quantizing it could push the first frames onto a wrong trajectory that decodes as a brief non-speech "music"/hum onset before the voice; keeping it full-precision avoids that at a small size cost.

Usage

Requires a voice reference audio (WAV file, 24 kHz mono) for voice cloning:

# Voice cloning TTS
VIBEVOICE_VOICE_AUDIO=reference_voice.wav \
crispasr --tts "Hello, how are you today?" \
    -m vibevoice-1.5b-tts-q4_k.gguf \
    --tts-output output.wav

Architecture

Single-LM architecture (differs from the streaming Realtime-0.5B):

• LM: Qwen2.5-1.5B (d=1536, 28 layers, 12 heads, 2 KV heads)
• Prediction head: 4 AdaLN + SwiGLU layers (d=1536)
• Acoustic encoder: 7-stage ConvNeXt (3200x downsample from 24kHz)
• Semantic encoder: same architecture, 128-dim latent
• Decoder: 7-stage transposed ConvNeXt (3200x upsample)
• DPM-Solver++: 20-step, cosine schedule, v-prediction

The model generates speech tokens autoregressively — the LM produces <|vision_pad|> (speech_diffusion) tokens that trigger diffusion sampling, with <|vision_start|> / <|vision_end|> as control tokens.

Quality

Input	Parakeet ASR
"Hello, how are you today?"	"Hello, how are you today?"

Differences from Realtime-0.5B

Feature	Realtime-0.5B	1.5B Base
Architecture	4L base + 20L TTS LM	Single 28L LM
Voice input	Pre-computed .pt prompts	Audio WAV files
Voice cloning	No (fixed presets)	Yes (from reference audio)
Multi-speaker	No	Yes (up to 4 speakers)
Streaming	Yes	No

License

MIT (same as original model).

Provenance and EU AI Act Art. 53 note

• Upstream model: microsoft/VibeVoice-1.5B — published by microsoft.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.