VibeVoice-ASR-Streaming-7B-DFlash2-Drafter

A DFlash 2 block drafter for VibeVoice-ASR-Streaming-7B (Ar4ikov/VibeVoice-ASR-Streaming-7B-AWQ-W4A16-ASYM), for speculative decoding in vibevoice.c. It proposes 8 tokens in one pass; vibevoice.c checks them in one pass of the model and keeps the ones the model agrees with, plus one of its own.

The check is exact: every checked row is computed with the arithmetic of the model's own one-token decode step, so the transcript with this drafter is byte-for-byte the transcript without it — words, timestamps, speakers. The drafter only changes how many passes it takes.

Use

Needs vibevoice.c with DFlash 2 support: branch dflash2 (PR #48), in the next release. The model and this drafter in one download: Ar4ikov/VibeVoice-ASR-Streaming-7B-AWQ-W4A16-ASYM-DFlash2 (the INT4 drafter in drafter/, used without --draft).

vv_cli --model ./VibeVoice-ASR-Streaming-7B-AWQ-W4A16-ASYM --audio talk.wav --draft ./VibeVoice-ASR-Streaming-7B-DFlash2-Drafter
vv_cli serve --model ./VibeVoice-ASR-Streaming-7B-AWQ-W4A16-ASYM --draft ./VibeVoice-ASR-Streaming-7B-DFlash2-Drafter --slots 4           # streaming sessions (WebSocket, SSE) too

It drafts for any VibeVoice-ASR-Streaming-7B checkpoint of the same model — microsoft/VibeVoice-ASR-Streaming-7B — since it reads only the model's hidden states, embedding and LM head; it was trained on the AWQ checkpoint's transcripts, so that is where it drafts best.

Results

vibevoice.c b72be15 (branch dflash2), RTX 3090, VibeVoice-ASR-Streaming-7B-AWQ-W4A16-ASYM, greedy decoding, decode tokens per second:

Streaming sessions (22 + 4 frames a chunk); plain = the same binary without --draft, decode tokens per second, --draft-check exact.

The drafter

• DFlash 2: 5 Qwen3-style layers (hidden 3584, 28/4 heads, intermediate 9472), 831M parameters, block 8.
• KV injection: for every position the model has processed, the outputs of its layers 1, 7, 13, 19 and 25 are concatenated, projected (fc) and normalized; every drafter layer turns them into keys and values. The drafter reads the model's own view of the audio and of the transcript so far.
• Two-tap dynamic convolution around attention and MLP (per-row kernels predicted from the row), and a candidate selector that re-ranks the model head's top 16 per row with a pairwise predecessor/successor term (rank 256).
• The embedding and the LM head are the model's, frozen. The drafter scores a draft vocabulary of 32768 ids (draft_vocab): the ids its training transcripts use, plus every stop id.
• Accepted drafts on 43 held-out clips' traces: 3.518 tokens per block of 8 (the selector's, with every earlier draft right).

Training

Self-distillation: the drafter learns what this model writes, timestamps and speaker ids included. tools/dflash in vibevoice.c holds the pipeline:

1. ~140 hours of audio, 2872 training clips and 43 held out, from LibriSpeech, FLEURS (8 languages), SOVA, AMI, earnings calls and VoxConverse — audio only; the datasets' own text is never used.
2. The model's own greedy transcripts of every clip (vv_dflash_data gen).
3. Traces: every position's token, its role and the five tapped layers' outputs, replayed through the runtime as streaming sessions (vv_dflash_data trace).
4. train.py (PyTorch, flex attention): anchors at generated positions, rows 1..7 scored against the next 7 tokens with weights e^(-k/4); cross-entropy of the head plus the selector's; AdamW, cosine schedule, BF16.

Limits

• vibevoice.c only: the checkpoint keeps the published DFlash 2 names, but the KV injection, the selector and the draft vocabulary are as vibevoice.c implements them (docs/DFLASH.md).
• CUDA only (the CPU path and Metal decode without the drafter). Greedy only.
• Audio the corpus covered little (Mandarin, Russian) keeps fewer drafts; the runtime then falls back to plain steps and stays near plain speed.

License

MIT, like VibeVoice.