shenava-fa-fastconformer-streaming-32m

Shenava Streaming 32M — Persian Cache-Aware Streaming ASR

A tiny (32M-param) cache-aware, multi-latency streaming Persian (Farsi) ASR model — FastConformer-Hybrid (RNNT + CTC), 16 kHz. Built for fully offline, on-device, real-time captioning (WebGPU / WASM / NeMo), part of the VisualEars project (SLT 2026).

One model serves the entire latency–accuracy curve (0 / 80 / 480 / 1040 ms) — pick your operating point at inference time, no re-training.

📊 Results — Golden6669 (held-out gold Persian eval)

Evaluated on Reza2kn/visualears-golden-6669 (6,669 clips, official Persian normalizer), RNNT head:

The curve is nearly flat — only 0.62 pp WER from 0 → 1040 ms. You get near-best accuracy at zero lookahead: a 32M model doing 11.08% WER / 3.17% CER at true real-time, fully on-device. For reference, the previous-generation single-latency fa32M scored 17.40% (and could not run below its trained latency).

Flatness holds per-condition (not just in aggregate)

The latency penalty is uniform across acoustic conditions — low-latency does not fray on the hard far-field/obstructed tail (Golden6669 is deliberately ~94% non-clean):

Far-field actually has the smallest 0→1040 ms gap. (Aside: clean scores worst here — a quirk of Golden6669's small 400-clip clean slice, identical across all latencies, not a streaming effect.)

WER_bf = boundary-forgiven WER (utterances perfect modulo Persian word-spacing conventions counted correct).

How it was trained

• Base: nvidia/stt_en_fastconformer_hybrid_medium_streaming_80ms (English cache-aware streaming), persianized by swapping in a Persian BPE-1024 tokenizer and reinitializing the decoder + joint (encoder kept).
• Multi-latency: trained with att_context_size = [[70,13],[70,6],[70,1],[70,0]] (chunked_limited) so a single checkpoint covers 0 / 80 / 480 / 1040 ms.
• Phase A: ~7,386 h / 3.66M clips of cleaned, teacher-pseudo-labeled Persian (Reza2kn/visualears-persian-pseudo-asr).
• Phase B: gold-anchor fine-tune on human-verified gold + active-learning corrections.
• Trajectory: random-init decoder → Phase A 17.94% → Phase B 10.46% (@1040 ms).
• Version: v2 (Phase B on 355 human-verified gold, vs v1's 300) — uniformly ~0.1 pp better WER at every latency (1040 ms 10.57→10.46%). Supersedes v1.

Usage (NeMo)

from nemo.collections.asr.models import ASRModel
m = ASRModel.restore_from("shenava-fa-streaming-32m.nemo").cuda().eval()
m.encoder.set_default_att_context_size([70, 0])      # 0 ms (real-time); or [70,13] for best WER
print(m.transcribe(["clip.wav"])[0].text)

[70,0]=0 ms · [70,1]=80 ms · [70,6]=480 ms · [70,13]=1040 ms (1 encoder frame = 80 ms, FastConformer subsampling 8).

Notes

• The companion offline flagship is Reza2kn/shenava-fa-fastconformer-115m (7.29% WER). A streaming 114M sibling (same recipe) is in training.
• The CTC head is the deployment head for real-time/WebGPU; RNNT is the higher-accuracy offline/rescorer head.

Citation

@misc{shenava_streaming_32m_2026,
  title  = {Shenava Streaming 32M: Persian Cache-Aware Streaming ASR},
  author = {Sayar, Reza},
  year   = {2026},
  howpublished = {Hugging Face},
  url    = {https://huggingface.co/Reza2kn/shenava-fa-fastconformer-streaming-32m}
}