---

license: apache-2.0

pipeline_tag: automatic-speech-recognition

language:

• zh

tags:

• audio
• speech-recognition
• chinese-dialects
• gguf
• ggml
• crispasr

base_model: DataoceanAI1/dolphin-cn-dialect-small-streaming

library_name: ggml

---

Dolphin CN-Dialect small (streaming) — GGUF

GGUF conversions of DataoceanAI1/dolphin-cn-dialect-small-streaming

for CrispASR (--backend dolphin, or -m dolphin to auto-download).

Dolphin CN-Dialect, from Dataocean AI and Tsinghua University, recognises Mandarin plus Chinese dialects. It is a

WeNet model: an E-Branchformer encoder, a Transformer attention decoder and a CTC head. CrispASR decodes it

the way upstream dolphin/transcribe.py does. It runs a CTC prefix beam search (beam 10) and rescores the n-best

with the attention decoder under the prompt [sos, <lang>, <region>, <asr>, <notimestamp>]. The decoder

predicts the language and region, or -l zh / -l zh-SICHUAN forces them.

Files

| File | Size | Notes |

| --- | --- | --- |

| dolphin-cn-dialect-small-streaming-f16.gguf | 792 MB | F16 matrices, F32 norms / convs / tables |

| dolphin-cn-dialect-small-streaming-q8_0.gguf | 443 MB | Q8_0 |

| dolphin-cn-dialect-small-streaming-q4_k.gguf | 258 MB | Q4_K (the default for -m dolphin) |

Verification

The F16 file was checked stage by stage against the upstream dolphin package (float32, CPU, dither 0) with

crispasr-diff dolphin. The checked stages are the fbank, the subsampling output, every encoder block, the encoder

output, the CTC log-probs and the final text. Every stage passed, and the transcripts are identical to upstream

on a Mandarin clip and on jfk.wav.

Quantised transcripts against F16 on 15 in-domain clips: Mandarin, Sichuan, Tianjin and Henan dialects,

Cantonese, and zh-en code-switching.

| | identical to F16 | differences |

| --- | --- | --- |

| Q8_0 | 13 / 15 | one dropped trailing 啊; one character (点 → 哪, Cantonese) |

| Q4_K | 13 / 15 | one dropped trailing 啊; one dropped trailing 的 |

Encoder-output cosines are less smooth than the transcripts suggest. A handful of frames sit near decision

boundaries, so the worst-frame cosine moves a lot under quantisation while the mean stays at ≥ 0.996 for Q8_0.

English (out of domain for this model) drifts more at Q4_K.

Licence

Apache-2.0, as upstream.