4,800
Concurrent live streams
On one H100 SXM 80 GB.
Models NVIDIA
Nemotron 3.5 ASR Streaming 0.6B
Multilingual speech recognition for live audio. With .wave, one H100 supports 4,800 concurrent streams at the 80 ms chunk setting—20× NVIDIA’s published concurrency.
$0.00045
Per minute of audio
Same model on Together AI: $0.0045/min.
40
Language locales
32 available directly; 8 require fine-tuning.
About the model
Nemotron 3.5 ASR is NVIDIA’s 600M-parameter model for streaming speech recognition. It turns incoming speech into punctuated text while the speaker is still talking, using one multilingual checkpoint.
Its cache-aware FastConformer encoder reuses earlier context instead of repeatedly processing overlapping audio. An RNN-T decoder generates the transcript. Audio chunks can be set to 80, 160, 320, 560 or 1,120 ms at inference time, without retraining.
On .wave, the model runs unchanged. Our benchmark supports 4,800 concurrent streams on one H100 SXM 80 GB at the 80 ms chunk setting.
Key capabilities
Native streaming
Transcribe new audio as it arrives, retaining the context from earlier chunks.
Configurable chunks
Adjust chunk duration to balance responsiveness and recognition accuracy. Chunk size is distinct from end-to-end latency.
Multilingual recognition
Provide the target locale, or let the model detect the spoken language automatically.
Formatted transcripts
Produce punctuation and capitalization as part of the model’s output.
Price comparison
| Service | Model | Per minute |
|---|---|---|
| .wave | Nemotron 3.5 ASR | $0.00045 |
| Together AI | Nemotron 3.5 ASR | $0.0045 |
| Together AI | Whisper Large v3 (Streaming) | $0.0035 |
Together AI rates checked 16 September 2026.
Applications & use cases
Multilingual voice agents
Provide live transcripts for customer support, in-car assistants and retail kiosks serving speakers of different languages.
Healthcare & clinical documentation
Turn medical dictation into text for clinical notes and documentation. Private deployment can support workflows that need to keep audio within a controlled environment.
Live captioning & transcription
Add captions to meetings, events and broadcasts as people speak. Make live conversations more accessible and easier to follow.
Post-call & offline analytics
Transcribe recorded calls for contact-center quality review, sales analysis and searchable conversation archives.
Languages
One checkpoint covers 40 locales. NVIDIA lists 32 for transcription out of the box and 8 that need fine-tuning. Regional variants include French (France and Canada), Portuguese (Brazil and Portugal), and Spanish (Spain and the US).
English en-US
English (UK) en-GB
Spanish es-ES- Spanish (US)
es-US 
French fr-FR
French (Canada) fr-CA
German de-DE
Italian it-IT
Portuguese (Brazil) pt-BR
Portuguese pt-PT
Dutch nl-NL
Polish pl-PL
Russian ru-RU
Ukrainian uk-UA
Czech cs-CZ
Slovak sk-SK
Hungarian hu-HU
Romanian ro-RO
Bulgarian bg-BG
Croatian hr-HR
Slovenian sl-SIFine-tuning
Greek el-GRFine-tuning
Swedish sv-SE
Danish da-DK
Finnish fi-FI
Norwegian (Bokmål) nb-NO- Norwegian (Nynorsk)
nn-NOFine-tuning 
Estonian et-EE
Latvian lv-LVFine-tuning
Lithuanian lt-LTFine-tuning
Maltese mt-MTFine-tuning
Turkish tr-TR- Arabic
ar-AR 
Hebrew he-ILFine-tuning
Hindi hi-IN
Chinese zh-CN
Japanese ja-JP
Korean ko-KR
Vietnamese vi-VN
Thai th-THFine-tuning