cabin-asr-vi-zipformer-offline-r4 (candidate, not a release)
Vietnamese offline (non-streaming) zipformer2 transducer, 68.6 M parameters, BPE-2000, for the Cabin Offline app. Same architecture and file names as the shipped sherpa-onnx-zipformer-vi-int8-2025-04-20. Fine-tuned from r3 rc1 on the same 258 h round-3 mix with additive noise (MUSAN) for robustness.
- Status: candidate, round 4 of the R13x → sherpa-vi distillation (trained 2026-10-30): r3 rc1 continued with additive-noise augmentation, better than r3 on real interpreter interviews (gate below). Drop-in for the shipped Vietnamese model (same files). Not a release candidate: the r3-style release check (int8 vs fp32, thread sweep), phone captures, the invented-card suite and ARM RTF are still to run; r3 rc1 stays the release candidate until then.
- Collection: Cabin offline ASR (on-device, distilled from R13x). Repo names:
doanhlt/cabin-asr-<lang>-<arch>-<variant>-<version>. - Code:
thinkprompt/cabin-offline-asr(private GitHub repo), branchfeat/mobile-local-kit,mobile-local/distill/+mobile-local/README.md("Pilot: R13x → sherpa-vi distillation"); round 4 =distill/run_xq_r4.sh(cab_xq queue,distill/xq_runner.sh), gatedistill/gate_r4.py, interviewseval/interviews/asr_pair.py.
Files
| path | what |
|---|---|
package/ |
the sherpa-onnx package the app loads (encoder / joiner int8, decoder fp32, tokens.txt, bpe.model) |
icefall/pretrained.pt |
icefall state dict: start point for the next fine-tune (--finetune-ckpt) |
eval/ |
paired-bootstrap comparisons (vs shipped, r3, the other arm), round-4 gate, interview pairs, publish gate, dev-pick logs |
data/yodas2_vi_heldout_*_clean.jsonl |
the two clean conversational held-out sets (refs) |
attribution/ |
YouTube video ids of all YODAS training audio (CC BY 3.0) |
Every file is listed with its sha256 in SHA256SUMS.txt (sha256sum -c SHA256SUMS.txt) and with its size in FILES.tsv.
Results (WER %, lower is better; 95 % paired bootstrap over utterances, 10,000 resamples; * = CI excludes 0)
| set (n) | shipped | r3 | this | this − shipped | this − r3 |
|---|---|---|---|---|---|
| FLEURS-100 vi (gate <= 10.22) (100) | 10.19 | 7.64 | 7.54 | -2.65 [-4.58, -1.24] * | -0.10 [-0.33, +0.10] n.s. |
| CV13-vi test200 (200) | 9.58 | 8.37 | 8.10 | -1.48 [-2.77, -0.20] * | -0.27 [-0.94, +0.35] n.s. |
| YODAS2 held-out clean, ref = R13x ∩ caption (463) | 7.43 | 5.12 | 5.05 | -2.38 [-2.81, -1.97] * | -0.07 [-0.23, +0.10] n.s. |
| YODAS2 held-out clean, ref = caption (899) | 52.59 | 51.62 | 51.60 | -0.98 [-1.34, -0.64] * | -0.02 [-0.17, +0.13] n.s. |
Round-4 gate (pre-registered before training, distill/gate_r4.py → eval/gate_r4.json): fleurs_le_10.22 = pass, not_worse_than_r3_tests = pass, beats_r3_interviews_all = pass, not_worse_than_r3_interviews_agree = pass, size_within_2pct_of_r3 = pass.
Interview units (real interpreter interviews, eval-only hold-out: 57 YouTube videos, R13x strict-hint reference, cluster bootstrap over videos; eval/asr_r4.json):
| subset | r3 | this | this − r3 | videos |
|---|---|---|---|---|
| all | 14.11 | 13.79 | -0.32 [-0.48, -0.16] | 57 |
| agree | 9.06 | 8.81 | -0.25 [-0.42, -0.07] | 52 |
| gemini | 46.26 | 45.50 | -0.76 [-1.56, -0.19] | 23 |
Publish gate (eval/gate_publish.json): fleurs100_le_10.22 = pass, sig_better_than_shipped_on_ge_2_sets = pass, not_sig_worse_than_previous_on_any_set = pass, size_within_5pct_of_shipped = pass, rtf_within_10pct_of_shipped = pass. Package 79.3 MB vs shipped 77.4 MB (same graph).
Attribution (r4n with noise − r4c control without noise): tests fleurs100_vi: +0.20 [-0.1, 0.53]; cv13_vi_test200: -0.27 [-1.28, 0.56]; yodas2_vi_heldout_agree_clean: -0.05 [-0.23, 0.13]; yodas2_vi_heldout_caption_clean: +0.05 [-0.09, 0.2]; interviews -0.26 [-0.49, -0.02].
Training data, licences and attribution
Same 258 h round-3 mix as r3 (identical cuts, same attribution files: YODAS / YODAS2 CC BY 3.0, FOSD CC BY 4.0, CV13 / CV17 CC0, FLEURS CC BY 4.0). Additive noise: MUSAN noise (CC BY 4.0), 2,211 10-s windows, mixed on the fbank features by lhotse CutMix (p 0.5, SNR 10-20 dB) during training only.
Licence: Apache-2.0 (base zzasdf/viet_iter3_pseudo_label Apache-2.0; teacher R13x, a Qwen3-ASR-1.7B fine-tune, Apache-2.0). No Gemini, Soniox or OpenAI output, no user capture, no eval set in training: FLEURS test/dev, CV13 test200, the YODAS2 dev / held-out videos and the interpreter-interview test videos are excluded (distill/check_holdout_cuts.py: 0 hits).
Recipe
- icefall zipformer
finetune.py --do-finetune 1from r3 rc1'spretrained.pt; LR 0.0005 (Eden lr_epochs 100, lr_batches 100k), 3 epochs, max-duration 200, fp16, one shared A100 (10 GB cap),--enable-musan 1(icefall CutMix, MUSAN noise p 0.5 SNR 10-20 dB). - Picked on dev only (FLEURS-vi dev + YODAS2-vi dev, mean WER) among (epoch, avg) 3:1 3:2 2:1: epoch 2 avg 1 (dev pick, runs/_logs/pick_r4n.log).
- Export: icefall
export-onnx.py, encoder / joiner int8, decoder fp32, non-causal (same as r3 / shipped).
How to use (sherpa-onnx, Python)
# pip install sherpa-onnx soundfile huggingface_hub
import sherpa_onnx, soundfile as sf
from huggingface_hub import snapshot_download
d = snapshot_download("doanhlt/cabin-asr-vi-zipformer-offline-r4", allow_patterns=["package/*"]) + "/package"
rec = sherpa_onnx.OfflineRecognizer.from_transducer(
encoder=f"{d}/encoder-epoch-12-avg-8.int8.onnx", decoder=f"{d}/decoder-epoch-12-avg-8.onnx",
joiner=f"{d}/joiner-epoch-12-avg-8.int8.onnx", tokens=f"{d}/tokens.txt",
num_threads=4, decoding_method="greedy_search") # greedy = what the app runs
audio, sr = sf.read("clip.wav", dtype="float32") # mono; sherpa-onnx resamples to 16 kHz
s = rec.create_stream(); s.accept_waveform(sr, audio); rec.decode_stream(s)
print(s.result.text)
The file names are those of the shipped sherpa-onnx-zipformer-vi-int8-2025-04-20, so the app swaps the model by
directory only.