Architecture: Cache-Aware Streaming FastConformer encoder (24 layers, d=1024, 8 heads) + RNN-T decoder (2-layer LSTM, hidden=640) + joint network (640 → 13088 vocab).

Languages: 39 languages, selected via prompt_kernel MLP conditioning:

Code	Language	Code	Language	Code	Language
ar-AR	Arabic	fr-CA	French (CA)	nn-NO	Norwegian (NN)
bg-BG	Bulgarian	fr-FR	French (FR)	pl-PL	Polish
cs-CZ	Czech	he-IL	Hebrew	pt-BR	Portuguese (BR)
da-DK	Danish	hi-IN	Hindi	pt-PT	Portuguese (PT)
de-DE	German	hr-HR	Croatian	ro-RO	Romanian
el-GR	Greek	hu-HU	Hungarian	ru-RU	Russian
en-GB	English (GB)	it-IT	Italian	sk-SK	Slovak
en-US	English (US)	ja-JP	Japanese	sl-SL	Slovenian
es-ES	Spanish (ES)	ko-KR	Korean	sv-SE	Swedish
es-US	Spanish (US)	lt-LT	Lithuanian	th-TH	Thai
et-EE	Estonian	lv-LV	Latvian	tr-TR	Turkish
fi-FI	Finnish	nb-NO	Norwegian (NB)	uk-UA	Ukrainian
		nl-NL	Dutch	vi-VN	Vietnamese
				zh-CN	Chinese

Key properties:

Sample rate: 16 kHz
128 mel filterbank features, n_fft=512, hop=160 (10ms), win=400 (25ms)
8× time downsampling (causal) → 80ms frame duration
Streaming: cache-aware attention with att_context_size=[[56,3],[56,0],[56,6],[56,13]]
Vocab: 13087 SentencePiece tokens + 1 blank (pure RNN-T, no TDT durations)
License: OpenMDW-1.1 (permissive)

Files

File	Size	Description
`nemotron-3.5-asr-streaming-0.6b-f16.gguf`	1.3 GB	F16 weights (full precision, F32 pre-encode)
`nemotron-3.5-asr-streaming-0.6b-q8_0.gguf`	686 MB	Q8_0 quantized (high precision, low footprint)
`nemotron-3.5-asr-streaming-0.6b-q4_k.gguf`	458 MB	Q4_K quantized (recommended, ~2x faster)

Usage with CrispASR

bash

# Download (Q4_K recommended — 458 MB, ~2x faster than F16)
huggingface-cli download cstr/nemotron-3.5-asr-streaming-GGUF \
  nemotron-3.5-asr-streaming-0.6b-q4_k.gguf --local-dir models/

# Transcribe (English, default)
crispasr --backend nemotron \
  -m models/nemotron-3.5-asr-streaming-0.6b-q4_k.gguf \
  -f audio.wav

# Transcribe in German
crispasr --backend nemotron \
  -m models/nemotron-3.5-asr-streaming-0.6b-q4_k.gguf \
  -f audio.wav --language de-DE

Conversion

Converted from the original NeMo .nemo checkpoint using:

bash

python models/convert-nemotron-to-gguf.py \
  --nemo nvidia/nemotron-3.5-asr-streaming-0.6b \
  --output nemotron-3.5-asr-streaming-0.6b-f16.gguf

Quantized variants can be produced with:

bash

crispasr-quantize models/nemotron-3.5-asr-streaming-0.6b-f16.gguf \
  models/nemotron-3.5-asr-streaming-0.6b-q4_k.gguf Q4_K

Architecture

Audio (16kHz) → Mel (128 bins, 10ms hop)
  → Pre-encode (3× causal Conv2d, 8× downsample, Linear 4352→1024)
  → 24× Cache-Aware FastConformer block:
      FFN1(½) → MHA(rel_pos, cache-aware) → DWConv(k=9, causal, LN) → FFN2(½) → LN
  → Prompt kernel (MLP: concat(enc[1024], lang_onehot[128]) → 2048 → 1024)
  → RNN-T decoder:
      Prediction: Embed(13088, 640) + 2-layer LSTM(640)
      Joint: enc(1024→640) + pred(640→640) → ReLU → Linear(640→13088)
  → Greedy / beam search decode

Original model

Paper: NVIDIA NeMo documentation
Source: nvidia/nemotron-3.5-asr-streaming-0.6b
License: OpenMDW-1.1

cstr/nemotron-3.5-asr-streaming-0.6b-GGUF

Nemotron-3.5-ASR-Streaming-0.6B GGUF

Model details

Files

Usage with CrispASR

Conversion

Architecture

Original model

cstr/
nemotron-3.5-asr-streaming-0.6b-GGUF