Nemotron 3.5 ASR Streaming 0.6B vs Voxtral Mini 4B Realtime 2602

Nemotron 3.5 ASR Streaming 0.6B

A small streaming transcription cousin for live audio in about 32 languages, with language detection and 80 ms to 1.12 s latency. It runs in under 8 GB in Transformers or as a 0.74 GB GGUF in NeMo-Speech.cpp, and several providers host it. The main gaps are no documented timestamps and weak accuracy outside the top 19 locales.

Voxtral Mini 4B Realtime 2602

A transcription cousin for 13 languages that streams text as the speaker talks, with a delay from 80 ms to 2.4 s, under Apache 2.0 and on Mistral's API. It needs a 16 GB GPU at BF16 through vLLM, Transformers or llama.cpp, or under 4 GB at Q4_K. The main gaps are no timestamps, safety results or answers from Mistral.

Nemotron 3.5 ASR Streaming 0.6B

1. Quickest local run, pip install "transformers>=5.13.0", then pipeline("automatic-speech-recognition", model="nvidia/nemotron-3.5-asr-streaming-0.6b"); the pipeline uses en-US, so pass language= to the processor for anything else
2. With NeMo-Speech.cpp, nemo-speech transcribe audio.wav --model nemotron-3.5-asr-streaming-0.6b.q8_0.gguf --language auto
3. Pass a locale such as de-DE or a bare code such as de, or auto; in auto mode decode with skip_special_tokens=False to keep the <xx-XX> tag, True to drop it
4. Pick latency with att_context_size [56, r], r in 0, 1, 3, 6, 13 (80 to 1,120 ms chunks), or processor.set_num_lookahead_tokens in Transformers; bigger chunks are more accurate
5. Use nvidia/nemotron-speech-streaming-en-0.6b instead for English only, as NVIDIA recommends
6. Prefer model.safetensors or the GGUF over the .nemo file, which is a pickle
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions

Voxtral Mini 4B Realtime 2602

1. Serve it with VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}' and install soxr, librosa and soundfile; mistral_common 1.9.0 or later comes with vLLM
2. Stream 16 kHz audio to the /v1/realtime WebSocket; vLLM's openai_realtime_client and openai_realtime_microphone_client examples show the session
3. Set temperature 0.0. Change the delay with transcription_delay_ms in tekken.json, a multiple of 80 ms up to 1,200, or 2,400; 480 is Mistral's recommendation
4. One text token is 80 ms, so set --max-model-len to at least seconds / 0.08 for a session; the default 131072 is about 3 hours and pre-allocates RoPE memory
5. In Transformers (5.2.0 or later), resample with processor.feature_extractor.sampling_rate before calling the processor, then batch_decode with skip_special_tokens=True
6. For a small machine, ggml-org's Q4_K_M GGUF with its mmproj runs in llama.cpp, and Mistral Experimental's Q4_K runs in audio.cpp
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions