NVIDIA Speech NIM (the current packaging of Riva ASR) runs as a Docker container on your GPU. You stream 16 kHz PCM over a WebSocket, get interim and final transcripts, and can turn on speaker diarization. There is no public per-minute rate: production use needs an NVIDIA AI Enterprise license priced per GPU, plus the GPU itself.
The LecSync Realtime Speech API is a managed WebSocket: preview and final captions, translation between any two of 60 languages, a speaker and audio-aligned timestamps on every final segment, and running minutes. It costs $0.006 per audio minute ($0.36 per hour) with those features included.
NVIDIA figures below were checked on October 23, 2026, against the ASR support matrix, the realtime WebSocket API, deploy and run ASR models, the AI Enterprise pricing page, and the NIM product FAQ.
The short answer
If you already run NVIDIA GPUs, need on-premises or air-gapped speech recognition, or must keep audio inside your own network, Speech NIM is the fit. Nemotron ASR Streaming can push thousands of concurrent streams on one large GPU, and AI Enterprise comes with enterprise support and defined SLAs.
If you want live translated captions, speakers, and minutes on one flat minute bill without standing up containers, LecSync is simpler and usually cheaper at lecture or meeting scale. 1,000 hours cost $360. A single AI Enterprise GPU subscription alone lists at $4,500 a year before you pay for the GPU.
How each API works
LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, a language pair and the features you want, then streams 16 kHz mono PCM to the URL it gets back. Preview text arrives while someone is still talking, then a final segment once the sentence settles. Translations come paired with the segment ID. With speakers: true, each final carries a speaker label and startMs/endMs on the recording's clock. With digest: true, minutes arrive section by section. The first word shows up about 0.28 seconds after speech starts, and translation adds roughly 88 ms.
NVIDIA Speech NIM (Riva ASR). You deploy a container such as nemotron-asr-streaming or a Parakeet model from NGC, with NIM_TAGS_SELECTOR choosing language profile and streaming mode. For the realtime HTTP/WebSocket path you POST to /v1/realtime/transcription_sessions on port 9000, then open ws://<host>:9000/v1/realtime?intent=transcription with PCM16 audio (default 16 kHz, mono). The server streams conversation.item.input_audio_transcription.delta events and a completed event with the full transcript; word times and speaker_tag appear when those options are on. gRPC on port 50051 remains available for classic Riva clients.
Two operational details matter in production:
- You own the GPU and the license. The NIM FAQ states production use requires NVIDIA AI Enterprise. Self-managed list price is $4,500 per GPU per year; public-cloud marketplace pricing is about $1 per GPU-hour plus the CSP instance. Prototyping on build.nvidia.com or the Developer Program is free and is not a production path.
- Idle sockets time out. The realtime server defaults to a 60-second inactivity timeout. LecSync caps server-side backlog at 2 seconds instead of dropping the session for short silences; see how LecSync caps backlog at 2 seconds.
Languages
Nemotron ASR Streaming ships an English-only profile and a multilingual profile covering 40 locales (19 listed as transcription-ready, 13 broad-coverage, 8 adaptation-ready). The multi profile can auto-detect with language_code=auto. Parakeet RNNT Multilingual covers 25+ languages with streaming; several Parakeet CTC models cover English, Spanish, Mandarin, Taiwanese and Vietnamese code-switching. Whisper Large v3 and Canary are offline-only in the current support matrix.
LecSync recognizes 60 languages, detects which one is being spoken, and translates between any two of them on the same stream.
Speakers and timestamps
On NIM, set speaker_diarization.enable_speaker_diarization and a max_speaker_count. Nemotron profiles use an eight-speaker Sortformer model; Parakeet CTC/RNNT profiles use a four-speaker model. The limit is an upper bound, not a promise that many speakers will appear. Completed events can include per-word start_time, end_time, confidence and speaker_tag when word time offsets are enabled.
LecSync attaches a speaker and millisecond timestamps to every final segment on one recording clock across reconnects. See building a meeting client with speakers and timestamps.
Translation and minutes
Streaming ASR NIM returns transcripts only. NVIDIA's own overview describes live translation as a pipeline: ASR NIM, then a separate NMT NIM, optionally TTS. Whisper Large v3 can translate to English offline; Canary 1b does bidirectional translation offline. Neither is a streaming any-to-any caption path on one socket.
LecSync returns translations on the same connection. For the DIY alternative, see built-in translation vs a DIY STT + MT pipeline. NIM has no live digest or minutes feature comparable to LecSync's digest: true; see live digest vs post-call summarization.
Side-by-side facts (checked October 23, 2026)
| LecSync Realtime Speech API | NVIDIA Speech NIM / Riva ASR | |
|---|---|---|
| Price model | $0.006/min ($0.36/hr), all features | No public $/min; AI Enterprise from $4,500/GPU/year or ~$1/GPU-hour + CSP |
| Free tier | Not listed on the developers page | build.nvidia.com + Developer Program for prototyping; 90-day AI Enterprise trial |
| Hosting | Managed SaaS, 7 gateway regions | Self-hosted Docker on your GPU (CC ≥ 8.0, ≥16 GB VRAM) |
| Streaming languages | 60, auto-detected | Nemotron multi 40 locales; Parakeet RNNT 25+; English-only Nemotron option |
| Live translation | Any two of 60, same socket | Not on streaming ASR; chain NMT NIM (or offline Whisper/Canary) |
| Interim results | Preview, then final | delta + completed events |
| Speakers | Every final segment, with startMs/endMs | Optional; max 8 (Nemotron) or 4 (Parakeet) |
| Live minutes | digest: true | No |
| Custom vocabulary | Glossary and domain context | Word boosting lists |
| Concurrency | Not published | Realtime max_connections default 10,000; Nemotron bs=512 up to 4,096 pipeline streams |
| Audio format | 16 kHz mono PCM (typical) | pcm16, 16 kHz default |
| Uptime commitment | Best effort, no SLA | AI Enterprise: defined SLAs and enterprise support |
What 1,000 hours cost
| Workload, 1,000 hours a month | LecSync | NVIDIA Speech NIM |
|---|---|---|
| Same-language captions | $360 | License + GPU (example floor: $4,500/GPU/year list before hardware) |
| Translated captions | $360 | ASR NIM + separate NMT NIM (second container / more GPU) |
| Speakers with timestamps | Included | Included when diarization enabled |
| Live minutes | Included | Not available |
1,000 hours is 60,000 minutes × $0.006. NVIDIA does not publish a per-minute ASR price, so you cannot line up a clean $/min without knowing how many streams share each GPU. At low to mid volume, the annual license alone already exceeds LecSync's $360. At very high concurrency on one large GPU, NIM's effective cost per minute can undercut a managed API — that is the honest trade. Other flat-vs-add-on vendors are in what 1,000 hours cost once translation, speakers and summaries are added.
When NVIDIA Speech NIM is the better pick
- On-premises or private cloud. Audio never leaves your network; you pick the GPU and region.
- Very high concurrency on owned GPUs. Nemotron's largest profile targets thousands of concurrent pipeline streams on a 48 GB-class GPU.
- Enterprise contract and SLA. AI Enterprise includes support and defined service-level agreements. LecSync has no SLA today.
- Customization depth. Word boosting, profile selection (low-latency vs high-throughput), and model choice per language.
- Already on the NVIDIA stack. If you already run other NIMs, adding ASR is another container, not another vendor.
When LecSync fits better
- Translated live captions. Any pair of 60 languages on the same socket, each translation tied to its segment.
- No GPU ops team. One
POST /v1/realtime/connect, seven managed gateway regions, no Docker or NGC keys. - Lectures and meetings with minutes. Speakers, timestamps and
digestduring the session. - Predictable minute billing. $0.36 an hour with translation and speakers included, billed to fractions of a second; empty connections are not charged.
- Mixed-language rooms. Language is detected per segment instead of fixed per container profile.
Comparing other managed APIs? See LecSync vs IBM Watson Speech to Text, LecSync vs Google Cloud Speech streaming and LecSync vs Azure Speech realtime translation.
FAQ
How much does NVIDIA Riva / Speech NIM cost?
As of October 23, 2026, NVIDIA does not list a public per-minute ASR rate. Production NIM requires NVIDIA AI Enterprise: $4,500 per GPU per year on the self-managed price list, or about $1 per GPU-hour on CSP marketplaces plus the instance cost. Prototyping on build.nvidia.com and the Developer Program is free; a 90-day AI Enterprise trial is available for evaluation.
Does NVIDIA Speech NIM translate speech in real time on one stream?
No. The streaming ASR WebSocket returns transcripts. NVIDIA documents live translation as a multi-NIM pipeline (ASR then NMT, optionally TTS). Whisper and Canary offer translation only in offline modes. LecSync returns any-to-any translation of 60 languages on the same connection.
How many speakers can NIM identify?
Up to eight on Nemotron ASR Streaming (Sortformer) and up to four on Parakeet CTC/RNNT profiles. max_speaker_count is an upper bound. LecSync labels every final segment with a speaker and millisecond timestamps.
Can I use Speech NIM without owning a GPU?
For production, you still need NVIDIA-accelerated infrastructure under an AI Enterprise license (your data center or a CSP GPU instance). For try-outs, build.nvidia.com hosts endpoints for prototyping. LecSync is fully managed and does not require you to run GPUs.
Bottom line
Choose NVIDIA Speech NIM when you need self-hosted streaming ASR at high concurrency, GPU-level control, or an enterprise SLA. Choose LecSync when you need translated captions, speakers and live minutes on one flat $0.006/min bill without operating containers. You can open a test stream from the developers page with one POST /v1/realtime/connect.