OpenASR/dolphin-cn-dialect-small

Dolphin CN-Dialect Small · OpenASR

Chinese multi-dialect speech recognition -- a WeNet E-Branchformer (CTC + attention) tuned for Sichuan and 22 regional dialects

Native speech-to-text in the OpenASR runtime — engineered for peak performance on CPU & GPU, no Python at inference time.

---

✨ Highlights

• 🀄 22 Chinese dialects — a WeNet E-Branchformer tuned for regional Mandarin, with a standout Sichuan (川话) strength across Wu, Cantonese, Minnan, Shanghainese and more
• 🎯 Dialect-first accuracy — a reported ~38% relative gain on dialect recognition and ~16% lower CER versus the base Dolphin, without giving up standard Mandarin
• 🧩 Joint CTC + attention — an E-Branchformer encoder with a Transformer decoder and CTC/attention rescoring that OpenASR runs bit-exact against its golden reference
• 🐬 Chinese-focused, char-level — a compact small.cn checkpoint over a character vocabulary, purpose-built for zh audio including heavy accents and code-mixed speech
• 🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull dolphin-cn-dialect-small:fp16

# 3. Transcribe
openasr transcribe audio.wav --model dolphin-cn-dialect-small

All builds for this model:

openasr pull dolphin-cn-dialect-small:fp16
openasr pull dolphin-cn-dialect-small:q8
openasr pull dolphin-cn-dialect-small:q4

📦 Available builds

Quant	File (.oasr)	Size	RAM peak	RTF · M1 CPU	RTF · M1 GPU	ΔCER vs fp16
fp16	dolphin-cn-dialect-small-fp16.oasr	860 MB	2.37 GB	0.32×	0.26×	0.0%
q8_0	dolphin-cn-dialect-small-q8_0.oasr	494 MB	2.88 GB	0.24×	0.12×	0.0%
q4_k	dolphin-cn-dialect-small-q4_k.oasr	298 MB	2.65 GB	0.26×	0.11×	4.5%

_{RTF = real-time factor on a 2.38s in-language Sichuan-dialect (川话) clip (lower is faster); RAM peak measured per pack in an isolated subprocess. ΔCER compares each quantized build's Sichuan-clip transcript to this model's fp16 Sichuan-clip transcript, so it measures quantization drift rather than absolute recognition accuracy. fp16 is the recommended default — near-reference quality at a fraction of the footprint.}

🧠 About Dolphin CN-Dialect Small

Dolphin CN-Dialect Small is a Chinese multi-dialect speech-recognition model from DataoceanAI, built on the Dolphin / WeNet recipe as an E-Branchformer encoder + Transformer decoder trained with a joint CTC + attention objective (the small.cn checkpoint over a character vocabulary). It specializes in regional Mandarin — a standout Sichuan (川话) capability alongside 22 Chinese dialects such as Wu, Cantonese, Minnan and Shanghainese — while keeping strong standard-Mandarin transcription (the card reports a large relative gain on dialect recognition and a meaningful CER reduction over the base Dolphin). This OpenASR repo repackages the weights as .oasr packs that run natively in the OpenASR runtime — no Python at inference, all decoding local. OpenASR decodes it with CTC beam search plus attention rescoring and verified the transcript bit-exact against a golden reference on a Sichuan-dialect clip. It ships in fp16 (maximum fidelity, recommended), q8_0, and q4_k builds.

⚙️ How these packs were made

Converted from DataoceanAI1/dolphi-cn-dialect-small with the OpenASR importer:

openasr model-pack import dolphin <src> <out>.oasr \
  --package-id dolphin-cn-dialect-small --quantization {fp16,q8-0,q4-k}

The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low.

⚖️ License

These packs inherit the upstream model's license: Apache-2.0 (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.

🙏 Acknowledgements

This pack is a redistribution of Dolphin CN-Dialect Small, created and open-sourced by DataoceanAI (DataoceanAI1/dolphi-cn-dialect-small). All credit for the original architecture, training, and weights belongs to the authors; the license is inherited from and identical to the upstream model (Apache-2.0). The model builds on the Dolphin multilingual ASR project and the WeNet E-Branchformer / joint CTC-attention recipe — thank you to the Dolphin and WeNet teams and to DataoceanAI for releasing their work openly. OpenASR only performs format conversion, quantization, runtime verification, and local-inference adaptation.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — DataoceanAI1/dolphi-cn-dialect-small