ONNX int8 exports of AI4Bharat's per-language IndicConformer hybrid CTC/RNNT checkpoints, prepared for on-device offline ASR through sherpa-onnx's OfflineRecognizer.from_nemo_ctc.

Each language directory carries export-meta.json with the sha256 of the source .nemo, so provenance is checkable rather than asserted.

Please cite AI4Bharat for the underlying models.

Eight languages: gu pa bn mr ml te ta kn.

<lang>/model.int8.onnx     140,337,395 B   (identical size across languages, DISTINCT weights)
<lang>/tokens.txt               67,605 B   5,633 entries
<lang>/export-meta.json                    source sha256, checksums, export settings

The eight files are byte-identical in size because they share an architecture. They are different models — the checksums above are the check that matters.

1. normalize_type must be per_feature. It is stamped on every graph. If it is stripped or set to "", the model loads, runs, and returns EMPTY with no error. Verify by decoding, never by a successful load.

Required ONNX metadata: vocab_size=5633, subsampling_factor=4, normalize_type=per_feature, feature_dim=80.

2. tokens.txt opens with Bengali tokens in every language directory — this is correct. The vocabulary is AI4Bharat's shared 22-language tokenizer (5,633 = 22 × 256 + 1) and the CTC head is multi-softmax. It looks like a packaging error and is not. The file is byte-identical across all eight (ee60967630213f31…); it is duplicated per directory so each bundle installs independently.

Unlike the 600m multilingual checkpoint, these per-language models did learn to suppress the other 21 blocks — measured 100% own-script output on all eight, 0/400 wrong-alphabet characters — so no language_id masking is required.

Scored on held-out ARTPARK-IISc/Vaani, 50 clips × 50 speakers per language, against the general-purpose multilingual Whisper bundle these replace.

Script purity is the headline, not WER. A 0% score means the incumbent was not producing the language's alphabet at all. Catastrophic failures fell from 93/400 to 5/400.

te and kn remain insertion-bound and are not recommended without further evaluation.

Conformer self-attention is O(T²) in sequence length, so memory is a function of how long the speaker talks. Measured in-process (Android arm64, sherpa-onnx 1.13.4, int8, 2 threads, one process per utterance, ±1 MB reproducible):

Budget accordingly: a typical 3–10 s dictation utterance costs under 350 MB, but a two-minute monologue costs 2.7 GB. If you have a memory ceiling, cap utterance length — that is the cheapest control and it belongs in your app, not in the model.

The int8 build quantizes all 54 Conv nodes to ConvInteger. Leaving them in fp32 — as some other re-exports do — makes ONNX Runtime allocate an im2col workspace that scales with clip length, pushing peak RSS to 2,099–2,635 MB from the same nominal precision. If you are getting 4× the memory you expect from an int8 Conformer, check your Conv node dtypes.

Cost of that choice: ConvInteger is ~3.6× slower than fp32 Conv on CPU.

Device-side accuracy, languages outside the eight, streaming/partial decoding (these are offline models), and the RNNT head (not exported — it needs numba/CUDA).