cstr/x-asr-zh-en-GGUF

X-ASR zh-en — GGUF

GGUF conversions of GilgameshWind/X-ASR-zh-en for CrispASR (--backend xasr, or -m xasr to auto-download).

X-ASR is an icefall streaming Zipformer2 transducer for Chinese and English with punctuation and casing, trained on about a million hours of speech. Upstream ships it as sherpa-onnx exports for four chunk sizes: 160, 480, 960 and 1920 ms. The four exports share every weight bit for bit, so one GGUF serves all of them, and CRISPASR_XASR_CHUNK_MS picks the chunk size (default 480).

CrispASR decodes exactly as sherpa-onnx's OnlineRecognizer greedy search does. That covers the same feature windows and caches carried from chunk to chunk, at most one symbol per frame, and the same text rendering. It also runs the model as a true streaming recogniser: the realtime WebSocket endpoint (crispasr --server --ws-port N, then ws://…:N+1/v1/realtime) emits append-only text as audio arrives.

Files

File	Size	Notes
x-asr-zh-en-f16.gguf	310 MB	F16 matrices, F32 convs / norms / tables
x-asr-zh-en-q8_0.gguf	168 MB	Q8_0 (the default for -m xasr)
x-asr-zh-en-q4_k.gguf	93 MB	Q4_K

Converted from the sherpa-onnx exports. The repository's streaming_exp/pretrained.pt is a different, earlier checkpoint: none of its tensors equal the exported weights, and it transcribes poorly.

Verification

The reference is the icefall PyTorch modules, loaded with the exported weights and driven chunk by chunk like sherpa-onnx. That reference reproduces onnxruntime's encoder output (cos ≥ 0.9999997) and sherpa-onnx's text exactly on both clips at 480 ms and 160 ms. Against it, crispasr-diff xasr on the F16 file:

stage	jfk (480 ms)	zh (480 ms)	jfk (160 ms)	zh (160 ms)
fbank, embed, 6 Zipformer stacks, encoder output: worst-frame cos	≥ 0.99999	≥ 0.99999	≥ 0.99999	≥ 0.99999
greedy tokens	identical	identical	identical	identical
text	identical	identical	identical	identical
streamed in 370 ms pieces	identical to one-shot	identical	identical	identical

Quantised transcripts: the Mandarin clip is identical at Q8_0 and Q4_K. On jfk.wav, Q8_0 gives "ask" for "asked", the same near-tie the 160 ms export shows. Q4_K also drops the English punctuation.

Licence

Apache-2.0, as upstream.