cabin-asr-vi-zipformer-streaming-r1 (candidate, not a release)
Vietnamese streaming (causal, chunked) zipformer2 transducer, BPE-2000 (tokens.txt / bpe.model of the shipped sherpa-vi), two packages: 640 ms chunks (package-c32/) and 320 ms chunks (package-c16/), left context 128 frames. Arm xqs: causal conversion of offline r2, 3 epochs on the round-3 mix, LR 0.0015. Round-3 mix = 258 h (r2 cuts + 168 h R13x-labelled YODAS v1 vi).
- Status: candidate (2026-10-30): first streaming Vietnamese model of the R13x distillation (cab_xq queue). Not in the app yet (the app decodes VAD segments offline); for live partials / low-latency captions.
- 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"); streaming =train/vi_zipformer_streaming/run_stream_xq.sh, selectiondistill/xq_follow.py.
Files
| path | what |
|---|---|
package-c32/ |
sherpa-onnx OnlineRecognizer package, 640 ms chunks (encoder / joiner int8, decoder fp32) |
package-c16/ |
same weights, 320 ms chunks |
icefall/pretrained.pt |
causal icefall state dict: start point for the next streaming fine-tune |
eval/ |
paired-bootstrap comparisons, gate, dev-pick log, selection record |
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)
| reference (640 ms) | fleurs100_vi | cv13_vi_test200 | yodas2_vi_heldout_agree_clean | yodas2_vi_heldout_caption_clean |
|---|---|---|---|---|
| this − streaming pilot (stream-c32) | 9.31 → 9.48 : +0.17 [-0.38, +0.73] n.s. | 11.74 → 10.73 : -1.01 [-2.49, +0.46] n.s. | 8.32 → 7.58 : -0.74 [-1.09, -0.41] * | 53.16 → 53.02 : -0.14 [-0.42, +0.13] n.s. |
| this − offline r2 rc1 | 7.44 → 9.48 : +2.04 [+1.29, +2.82] * | 8.37 → 10.73 : +2.36 [+0.85, +3.93] * | 5.61 → 7.58 : +1.96 [+1.53, +2.47] * | 51.78 → 53.02 : +1.25 [+0.91, +1.60] * |
| this − shipped offline sherpa-vi | 10.19 → 9.48 : -0.71 [-2.58, +0.73] n.s. | 9.58 → 10.73 : +1.15 [-0.62, +2.82] n.s. | 7.43 → 7.58 : +0.14 [-0.36, +0.66] n.s. | 52.59 → 53.02 : +0.44 [+0.08, +0.78] * |
| 320 ms: this − stream-c16 | 11.28 → 10.39 : -0.88 [-1.81, +0.00] n.s. | 15.65 → 13.50 : -2.16 [-3.97, -0.53] * | 9.80 → 8.55 : -1.24 [-1.73, -0.75] * | 53.86 → 53.40 : -0.46 [-0.80, -0.14] * |
Selection (pre-registered, DECISIONS.log 2026-10-30): the winner is the arm with the lowest mean dev WER (FLEURS-vi dev + YODAS2-vi dev, 640 ms) among the arms whose own pre-registered test gate passed (xqs: vs stream-c32; xqs4 / xqs3i: vs xqs). Test sets were not used to rank.
| arm | recipe | mean dev WER (640 ms) | gate |
|---|---|---|---|
| xqs | causal conversion of offline r2, 3 epochs on the round-3 mix, LR 0.0015 | 6.072 | PASS |
| xqs4 | xqs continued 3 more epochs on the same mix, LR 0.001 (6 epochs in total) | 6.062 | FAIL |
| xqs3i | causal conversion of offline r3, 3 epochs on the round-3 mix, LR 0.0015 | 6.015 | FAIL |
Winner: xqs. The lowest dev WER (6.015) belongs to an arm whose test gate did not pass, so it was not eligible.
Training data, licences and attribution
Round-3 mix of cabin-asr-vi-zipformer-offline-r3 (same cuts, same attribution: YODAS / YODAS2 CC BY 3.0, FOSD CC BY 4.0, CV13 / CV17 CC0, FLEURS CC BY 4.0, MUSAN CC BY 4.0, ESC-10 CC BY). Video ids: see that repo's attribution/.
Licence: Apache-2.0 (base zzasdf/viet_iter3_pseudo_label Apache-2.0; teacher R13x Apache-2.0). No Gemini, Soniox or OpenAI output, no user capture, no eval set in training (FLEURS test/dev, CV13 test200, YODAS2 dev / held-out and the interpreter-interview test videos are excluded).
Recipe
train/vi_zipformer_streaming/run_stream_xq.sh(icefall zipformer, causal, dynamic chunks 16,32,64,-1, left 64,128,256,-1), max-duration 200, fp16, one shared A100 (6 GB cap).- Dev pick (epoch, avg) in 3:1 3:2 2:1 at 640 ms; export
export-onnx-streaming.pyat chunk 16 / 32, left 128, encoder / joiner int8.
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-streaming-r1", allow_patterns=["package-c32/*"]) + "/package-c32"
rec = sherpa_onnx.OnlineRecognizer.from_transducer(
encoder=f"{d}/encoder.int8.onnx", decoder=f"{d}/decoder.onnx", joiner=f"{d}/joiner.int8.onnx",
tokens=f"{d}/tokens.txt", num_threads=2, sample_rate=16000, feature_dim=80, decoding_method="greedy_search")
audio, sr = sf.read("clip.wav", dtype="float32")
s = rec.create_stream()
for i in range(0, len(audio), 1600): # 100 ms pieces, like a microphone callback
s.accept_waveform(sr, audio[i:i + 1600])
while rec.is_ready(s):
rec.decode_stream(s)
print(rec.get_result(s)) # live partial
s.input_finished()
while rec.is_ready(s):
rec.decode_stream(s)
print(rec.get_result(s))