NVIDIA: Nemotron 3.5 ASR Streaming Multilingual 0.6B

Nemotron 3.5 ASR Streaming Multilingual 0.6B is NVIDIA’s 600-million-parameter speech-recognition model for live transcription and recorded-audio workloads. It extends the English streaming line with language-ID conditioning. Speech becomes text in the input language, with punctuation and capitalization built into the output.

The maker’s language coverage has three tiers. Nineteen transcription-ready locales cover US and UK English, US and Spain Spanish, France and Canada French, Italian, Brazil and Portugal Portuguese, Dutch, German, Turkish, Russian, Arabic, Hindi, Japanese, Korean, Vietnamese and Ukrainian. Thirteen broad-coverage locales add Polish, Swedish, Czech, Norwegian Bokmål, Danish, Bulgarian, Finnish, Croatian, Slovak, Mandarin, Hungarian, Romanian and Estonian. Together these 32 locales produce transcription without additional fine-tuning.

Eight adaptation-ready locales—Greek, Lithuanian, Latvian, Maltese, Slovenian, Hebrew, Thai and Norwegian Nynorsk—are recognized by the tokenizer but require in-domain fine-tuning to enable full transcription. Thus the producer’s 40-locale figure is not a claim that all forty have the same out-of-the-box capability. A customized checkpoint is a separate artifact from the base model.

The architecture combines a 24-layer cache-aware FastConformer encoder with an RNN-T decoder. It retains encoder attention and convolution states from earlier frames and processes each new audio chunk without repeatedly recomputing overlapping windows. A language-ID signal is supplied alongside the acoustic representation to condition the transcript or let the model infer the language.

NVIDIA’s NeMo configuration documents chunk sizes of 80, 160, 320, 560 and 1120 milliseconds. The attention setting controls how much future audio a chunk receives before text is emitted, allowing a task to balance waiting time and acoustic context. These are producer streaming configurations, not guaranteed end-to-end response times; an implementation can expose its own supported subset.

Prepare mono WAV audio at the producer processor’s 16 kHz sample rate. When the language is known, use the corresponding locale conditioning; automatic mode detects the spoken language. The basic pipeline example uses a US-English prompt by default, so multilingual behavior should be configured deliberately rather than assumed from that example.

In automatic mode, the producer’s workflow can append a language tag after the transcript’s terminal punctuation. Keeping the tag makes the language available to an application handling mixed-language traffic; stripping it returns clean transcript text. Preserve the distinction between the recognized words, their punctuation and this language metadata when displaying or processing the result.

A captioning application can feed audio in order, collect the transcript and use completed text for its display or downstream processing. A voice application can pass recognized speech onward to its language-understanding stage. Call-analysis work can retain the original recording and transcript together so names, numbers and unfamiliar terms remain inspectable. The ASR model supplies the transcript; summarization, actions and speech generation are separate stages of the application.

NVIDIA trains the model on a blend of public and proprietary speech data with human and synthetic labels normalized to punctuated, properly cased text. Its adaptation guide emphasizes matching language labels and text style, then assessing held-out audio at the streaming setting intended for use. It describes language, accent and domain adaptation without making those customized results properties of the original checkpoint.

The model card ties practical recording length to available memory rather than publishing one universal duration ceiling. Keep the multilingual edition distinct from the English-only model, and retain the selected language and streaming configuration with the transcription result. The output artifact is the recognized text, with optional producer language metadata, rather than a general conversational answer.