sherpa-vi-distill-pilot1 (candidate, not a release)
Shipped app model sherpa-onnx-zipformer-vi-int8-2025-04-20 (zipformer2 transducer 68.6 M, Apache-2.0,
base zzasdf/viet_iter3_pseudo_label) fine-tuned on transcripts written by R13x (R13-1.7B RC1, Apache-2.0).
Same architecture, BPE-2000 and file names as the shipped package: the app swaps it by directory.
- Status: candidate, 2026-10-23. It passes every gate measured on the GPU box. It is NOT yet measured on the user's phone captures (private, Mac only) or the app's invented-card suite, which are the real gates.
- Code:
cabin-offline-asrbranchfeat/mobile-local-kit,mobile-local/distill/. Full write-up is inmobile-local/README.md, section "Pilot: R13x → sherpa-vi distillation". - Collection: Cabin offline ASR (on-device, distilled from R13x). New repos are named
doanhlt/cabin-asr-<lang>-<arch>-<variant>-<version>; this repo keeps its first name because links to it exist. Next version: cabin-asr-vi-zipformer-offline-r2.
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 (epoch 4, average over epochs 1–4): start point for the next fine-tune (--finetune-ckpt) |
eval/ |
per-set WER (summary.json), paired-bootstrap comparisons, dev-pick logs, mix stats |
data/distill_yodas_train.jsonl |
the 10,186 training rows (YODAS2 vi000 clip ids + R13x label + caption + agreement) |
data/yodas2_vi_heldout_*.jsonl |
the two conversational held-out sets (103 videos, disjoint from train) |
attribution/yodas2_video_ids.txt |
the 408 YouTube video ids the training audio comes from (CC BY 3.0) |
DECISIONS.log |
deviations from the plan and the reasons for them |
Results (WER %, 95 % paired bootstrap)
| set (n) | shipped | replay-only control | this | R13x teacher | this − shipped |
|---|---|---|---|---|---|
| FLEURS-100 vi (100), gate ≤ 10.22 | 10.19 | 8.73 | 7.91 | 5.74 | −2.28 [−4.09, −1.02] |
| CV13-vi test200 (200) | 9.58 | 9.65 | 9.11 | 10.19 | −0.47 [−1.56, +0.56] n.s. |
| YODAS2 held-out, ref = R13x ∩ caption (1,125) | 7.67 | 7.69 | 4.59 | – | −3.07 [−3.36, −2.79] |
| YODAS2 held-out, ref = caption (1,660) | 38.65 | 38.67 | 37.33 | 38.23 | −1.32 [−1.54, −1.10] |
- RTF on CPU, 4 threads, is unchanged: 0.023.
- The replay-only control (FLEURS train only, same protocol) does not move either YODAS set. The conversational gain therefore comes from the R13x labels.
- Most of the FLEURS gain comes from the FLEURS-train replay, not from distillation: the difference between this model and the control there is not significant.
Training data and licence
- Audio:
- YODAS2
vi000shards 400–449 (espnet/yodas2, CC BY 3.0, attribution list inattribution/): 27.65 h kept; - FLEURS vi train (CC BY 4.0) ×3 replay: 13.66 h.
- YODAS2
- Labels: R13x (Apache-2.0) hypotheses, kept where they agree with the YouTube caption (interior WER ≤ 0.30, length ratio 0.6–1.6, Vietnamese, no digits).
- Not used: Gemini, Soniox or OpenAI outputs, user captures, and any eval set.
- Licence: derivative of an Apache-2.0 model. Credit YODAS (CC BY 3.0) and FLEURS (CC BY 4.0) in the app's licences screen.
Recipe
- icefall zipformer
finetune.pywithdo-finetunefrom the shipped base: LR 0.0015, Eden lr_epochs 100 / lr_batches 100k, max-duration 200, fp16. - 5 epochs on one A100 shared with other jobs, 10 GB cap.
- Picked on dev only (FLEURS dev + YODAS2 dev): epoch 4, avg 3.
- One seed.
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-sherpa-vi-distill-pilot1", 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.