Nemotron Streaming

Two NVIDIA streaming ASR models share the NemotronStreamingASR Swift target. Both are 600M-parameter cache-aware FastConformer encoders paired with an RNN-T decoder, both emit native punctuation and capitalization as regular BPE tokens, both run on the Apple Neural Engine via CoreML, and the multilingual variant additionally has MLX bundles for GPU-resident inference. Pick the one that matches your application:

Shared properties

• Native punctuation and capitalization — no post-processor needed; periods, commas, and case are part of the vocabulary
• 600M parameters — cache-aware 24-layer FastConformer encoder + 2-layer LSTM RNN-T decoder + joint network
• Cache-aware streaming — attention and conv caches flow chunk-to-chunk for continuous context across chunk boundaries
• Multiple chunk sizes — multilingual: 80, 320, 560, 1120 ms (320 ms default); English: 80, 160, 560, 1120 ms (160 ms default)
• Multilingual-only: a prompt kernel (Linear 1152→2048→1024) folds a one-hot language slot into every encoded frame, so the same weights serve all 40 language-locales. The English-only variant has no prompt kernel.

Architecture

Three CoreML models pipelined per audio chunk:

No EOU head

Unlike Parakeet-EOU, Nemotron does not emit a dedicated end-of-utterance token. Two ways to segment continuous audio into utterances:

1. External VAD — pair the session with Silero VAD; on sustained silence, call finalize() to commit the current utterance and createSession() for the next one.
2. Punctuation boundary — when the partial transcript ends in ., ?, or !, treat that as a natural commit cue. No extra model, but depends on the audio actually inducing terminal punctuation.

Bundles

Four published variants of Nemotron-3.5-ASR-Streaming-0.6B, plus the older English-only model on the same Swift target:

Upstream: nvidia/nemotron-3.5-asr-streaming-0.6b (multilingual) and nvidia/nemotron-speech-streaming-en-0.6b (English-only).

Quantization is essentially lossless: CoreML INT8, MLX bf16, and MLX 8-bit are within ±0.3 pp WER of the fp32 NeMo source. MLX 4-bit costs ~6 pp avg WER for the smallest disk and streaming RSS.

Quick start — batch transcription

Conforms to SpeechRecognitionModel, so it drops into any code path that takes a generic STT model:

import NemotronStreamingASR

let model = try await NemotronStreamingASRModel.fromPretrained()
let text = try model.transcribeAudio(audioSamples, sampleRate: 16000, language: "en-US")

Quick start — async streaming

for await partial in model.transcribeStream(audio: samples, sampleRate: 16000, language: "ja-JP") {
    if partial.isFinal { print("FINAL: \(partial.text)") }
    else               { print("... \(partial.text)") }
}

Each PartialTranscript carries text, isFinal (true only for the last partial after finalize()), confidence, and a monotonic segmentIndex.

Long-lived session API (mic input)

let session = try model.createSession(language: "en-US")

// each mic chunk:
let partials = try session.pushAudio(float32Chunk16kHz)
for p in partials { showPartial(p.text) }   // isFinal is false mid-stream

// when the utterance ends (VAD silence or explicit stop):
let trailing = try session.finalize()
for p in trailing { commit(p.text) }

CLI

speech transcribe recording.wav --engine nemotron --language en-US                    # batch
speech transcribe recording.wav --engine nemotron --language en-US --stream           # streaming final
speech transcribe recording.wav --engine nemotron --language ja-JP --stream --partial # partials, Japanese
speech transcribe meeting.wav   --engine nemotron --language de-DE                    # any of the 40 locales

Nemotron vs Parakeet-EOU