A low-latency streaming Swahili ASR model: NVIDIA's cache-aware FastConformer-RNNT (nvidia/nemotron-3.5-asr-streaming-0.6b, 0.6B) fine-tuned to add Swahili, a language the base model does not officially support. Unlike Whisper (a 30-second-window batch model), this model transcribes incrementally with sub-100 ms latency — for live captioning, voice agents, and dictation.

This checkpoint is a weight-average ("soup") of the 5 best fine-tuning checkpoints, each selected on the held-out FLEURS Swahili test set (not the training-time dev metric).

Results — FLEURS `sw_ke` test (n=487, identical streaming pipeline, lowercase+punct-strip)

attention context (lookahead)	WER	CER
`[56,13]` (~480 ms)	0.298	0.106
`[56,0]` (~0 ms, lowest latency)	0.322	0.115
base model `[56,13]` (no Swahili FT)	1.147	0.860

The base model is effectively unusable for Swahili (~100% WER, degenerate ⁇ loops); fine-tuning brings it to ~30% WER / ~11% CER — within the range NVIDIA reports for its officially supported languages. For reference, offline (non-streaming) Whisper-large-v3 (1.5B) on the same test set is in a similar WER range but cannot stream.

Intended use & limitations

Good for: streaming/live Swahili transcription, voice agents (with an LLM downstream), live captioning, dictation aids, gisting. Not for: unattended verbatim transcription (legal/medical) without review.

Limitations:

Training data is formal, image-prompt-elicited standard ("Sanifu") Kenyan Swahili with noun-level English code-switching — not Sheng / spontaneous speech. Expect higher WER on Sheng-heavy or noisy spontaneous audio. FLEURS (read news) is a relatively favorable test.
0.6B streaming model: offline larger models are more accurate; the value here is latency.
Pick att_context_size for your latency/accuracy budget: more lookahead ([56,13]) is more accurate; [56,0] is lowest latency.

Usage (NeMo)

python

import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.restore_from("nemotron-3.5-swahili-streaming-0.6b.nemo")
# Streaming inference: use NeMo's
# examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py
# with target_lang=sw-KE, att_context_size=[56,13] (or [56,0]), decoder_type=rnnt.

Training

Base: nvidia/nemotron-3.5-asr-streaming-0.6b (FastConformer-RNNT, cache-aware streaming).
Data: ~382 h Kenyan Swahili from badrex/swahili-speech-400hr (Afrivoice / DigitalUmuganda, CC-BY-4.0); speaker-disjoint dev split. A 30 h clean subset is at Tonykip/kenyan-swahili-asr-clean.
Recipe: full fine-tune (all 637M params), AdamW, bf16, Noam schedule (warmup 2000), fused RNNT joint; pretrained tokenizer reused (vocab 13087). Best checkpoint = top-5 weight-average, selected on FLEURS.

ONNX / edge deployment

ONNX exports are in onnx/ for CPU / desktop / edge streaming via ONNX Runtime (Rust ort crate; parakeet-rs implements the streaming chunk + RNNT greedy loop):

encoder-swahili.int8.onnx (+_data, ~650 MB) — int8 encoder; the macOS/edge build (4× smaller than fp32)
encoder-swahili.onnx (+_data, ~2.4 GB) — fp32 master
decoder_joint-swahili.onnx — RNNT predictor + joint (keep fp32)
prompt_kernel.pt + prompt_info.json — the Swahili language prompt (sw-KE:48), applied host-side (the exported encoder is prompt-less)
onnx/INTEGRATION.md — cache I/O + prompt + streaming-loop spec

Graphs are NeMo check_trace-verified (ONNX≈PyTorch). End-to-end streaming WER is validated in the runtime. GGUF is not applicable (ggml/whisper.cpp is Whisper-only) — ONNX Runtime is the CPU/edge path.

License & attribution

Derivative of nvidia/nemotron-3.5-asr-streaming-0.6b under OpenMDW-1.1 — see the base model for license terms. Training data: badrex/swahili-speech-400hr (CC-BY-4.0, Afrivoice/DigitalUmuganda). Please attribute both NVIDIA and the data authors.

Tonykip/nemotron-3.5-swahili-streaming-asr

Nemotron 3.5 ASR — Swahili (streaming)

Results — FLEURS sw_ke test (n=487, identical streaming pipeline, lowercase+punct-strip)