# NVIDIA Riva

> Source: https://aiwiki.ai/wiki/nvidia_riva
> Updated: 2026-08-01
> Fact-checked: 2026-08-01
> Categories: AI Inference, Developer Tools, Enterprise AI, NVIDIA, Speech & Audio AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Riva." aiwiki.ai, 1 Aug 2026. https://aiwiki.ai/wiki/nvidia_riva
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA Riva** is a GPU-accelerated software development kit and family of containerized inference services for speech and translation AI, built by [NVIDIA](https://aiwiki.ai/wiki/nvidia). Riva serves [automatic speech recognition](https://aiwiki.ai/wiki/speech_recognition) (ASR), [text to speech](https://aiwiki.ai/wiki/text_to_speech_ai) (TTS), [neural machine translation](https://aiwiki.ai/wiki/neural_machine_translation) (NMT), speaker diarization, voice activity detection and punctuation restoration through a single gRPC API, and it is designed to be deployed by the customer rather than consumed as a hosted cloud service. NVIDIA describes it as "a GPU-accelerated SDK for building Speech AI applications that are customized for your use case and deliver real-time performance," running on premises, in any cloud, at the edge and on embedded hardware.[1]

Riva occupies a specific slot in NVIDIA's software stack. The [NeMo](https://aiwiki.ai/wiki/nvidia_nemo) toolkit trains and fine-tunes the models; Riva takes the resulting checkpoints, compiles them with [TensorRT](https://aiwiki.ai/wiki/tensorrt), and serves them on [Triton Inference Server](https://aiwiki.ai/wiki/nvidia_triton_inference_server) behind a speech-specific API.[1][10] The models it serves are largely the same ones NVIDIA publishes openly, including the [Parakeet](https://aiwiki.ai/wiki/parakeet) and [Canary](https://aiwiki.ai/wiki/canary) families, but the Riva server itself is proprietary software licensed under NVIDIA's terms rather than an open-source project.[24]

The product has moved twice. It began life as NVIDIA Jarvis and was renamed Riva on July 28, 2021, with NVIDIA stating that "the core technologies, performance and roadmap remain unchanged."[12] More recently the packaging inverted: as of release 2.24.0 NVIDIA deprecated x86 data center deployment of the Riva SDK entirely and now directs data center users to [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) microservices for ASR, TTS and NMT, leaving the standalone Riva SDK release supporting only embedded L4T platforms.[3][4] At the same time NVIDIA rebranded the underlying model family as "Nemotron Speech," describing it on the Riva developer page as "a family of GPU-accelerated open models" that are "deployed through the NVIDIA Riva library for optimized inference."[11] Riva today is therefore best understood as the speech inference runtime inside NVIDIA's speech products rather than as a standalone product a data center customer installs.

## From Jarvis to Riva

NVIDIA developed the technology under the Jarvis name and released it through a beta program. The rename to Riva was announced on the NVIDIA developer forums on July 28, 2021; all references to Jarvis in APIs, NGC containers and developer resources were replaced, older Jarvis-named applications continued to work but were deprecated, and the technology itself was unchanged.[12] Traces of the old name survive in the release notes, where the 1.3.0 Beta upgrade instructions still tell users to rerun `jarvis-build` and to run `jarvis_clean.sh` followed by `jarvis_init.sh`.[3]

At GTC in November 2021, NVIDIA announced Riva Enterprise, a paid tier that added NVIDIA expert support for large-scale deployments while Riva remained "available for free to customers and partners with smaller workloads."[13] That announcement also introduced custom voice creation, which NVIDIA said allowed an organization to build "a new neural voice with 30 mins of audio data in a day on A100."[13]

Riva 2.0 reached general availability at GTC in March 2022.[14] The 2.0.0 release notes record the break with the past: "All legacy Jarvis APIs have been removed and are no longer supported." The same release added Linux ARM64 support for Jetson AGX Xavier and Jetson NX Xavier, phoneme SSML tags for manual pronunciation overrides, and SSL/TLS connections to the Riva server.[3] NVIDIA said the Riva Enterprise tier covered "unlimited use of ASR and TTS services on any cloud and on-prem platforms" plus expert support and priority access to new releases.[14]

By September 2022 NVIDIA was shipping pretrained models in seven languages, having added French and Hindi, with word-level confidence scores and speaker diarization as new capabilities.[15] The standalone Riva and Riva Enterprise product pages on nvidia.com now redirect to the developer topic page, and access runs through [NVIDIA AI Enterprise](https://aiwiki.ai/wiki/nvidia_ai_enterprise) instead.[11][25]

### Release history

The public release notes document the following, among others. Note that the published notes jump from 2.19.0 to 2.24.0, so intermediate 2.2x releases are not separately documented there.[3]

| Release | Triton Inference Server | TensorRT | Notable changes |
| --- | --- | --- | --- |
| 1.10.0 Beta | 2.19.0 | 8.2 | Request-time ASR word boosting introduced |
| 2.0.0 (March 2022) | not stated | not stated | First GA release; all legacy Jarvis APIs removed; Jetson AGX Xavier and NX Xavier support; SSL/TLS |
| 2.17.0 | 2.47.0 | 10.1 | Offline Parakeet-RNNT, Whisper and Distil-Whisper support; automatic speech translation using Whisper |
| 2.18.0 | 2.50.0 | 10.4 | Canary ASR architecture in offline mode; T5 TTS; HTTP/REST for offline ASR; unified server and service-maker container |
| 2.19.0 | 2.54.0 | 10.8 | Streaming speaker diarization via Sortformer; Magpie-TTS multilingual; Silero VAD endpointing; Volta V100 and earlier deprecated |
| 2.24.0 | 2.60.0 | 10.13 | x86 data center deployments deprecated in favor of Riva NIM; Jetson Orin deprecated in favor of Jetson Thor; NVIDIA Fleet Command support dropped; FP8 quantization for Parakeet-CTC English |
| 2.26.0 | 2.60.0 | 10.13.2.6 | Nemotron-ASR English streaming model on [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor) with bundled Sortformer diarizer; NMT on Jetson via Megatron 1B any-to-any; WebSocket realtime API on port 9000 and HTTP REST on 9001 alongside gRPC on 50051 |

Release 2.26.0 is the current Riva Skills version in the documentation, and the `nvcr.io/nvidia/riva/riva-speech:2.26.0-l4t-aarch64` container was published to the NGC catalog in June 2026.[2][21] Its support matrix lists only embedded aarch64: Jetson Thor on JetPack 7.x, with CUDA 13.0.48, cuDNN 9.12.0, TensorRT 10.13.2.6 and Triton Inference Server 2.60.0 inside the container.[4]

## How Riva is built and deployed

Riva does not run NeMo checkpoints directly. NVIDIA's documentation is explicit that "the Riva server cannot use unsupported NVIDIA NeMo models directly because they represent only a single model," whereas Riva serves whole pipelines that include pre- and post-processing components.[10] Getting a model into Riva is a three-phase process:[10]

1. **Development.** Models are trained or fine-tuned in NeMo and saved as `.nemo` files, which are converted to Riva's `.riva` format. Newer releases accept `.nemo` files directly in `riva-build` through integrated `nemo2riva` conversion, making the separate conversion step optional.
2. **Build.** The `riva-build` tool gathers models, configuration files and user settings into an intermediate artifact called an RMIR (Riva Model Intermediate Representation). Supported pipelines are `speech_recognition`, `speech_synthesis`, `punctuation` and `translation`.
3. **Deploy.** The `riva-deploy` step converts the RMIR into a Riva model repository, exporting and optimizing the networks for the specific target hardware. This must run on the machine or cluster that will serve the model.

The result is served by Triton Inference Server behind Riva's gRPC API. NVIDIA states that "Riva's high-performance inference is powered by NVIDIA TensorRT optimizations and served using the NVIDIA Triton Inference Server," and that Riva services are exposed "as gRPC-based microservices for low-latency streaming, as well as high-throughput offline use cases."[1] Deployment to Kubernetes clusters is handled with Helm charts.[1]

Because the pipelines are compiled for a specific target, a Riva build is not portable between GPU architectures without redeployment. That tight coupling is what makes the containerized NIM packaging attractive: the container ships pre-compiled profiles instead.

## Services and capabilities

### Speech recognition

Riva ASR runs in offline (batch) and streaming modes. In offline mode the full audio is captured before the request is made, which NVIDIA notes "can have long latency"; in streaming mode the client sends chunks and the server returns intermediate transcripts as soon as they are available.[5]

The documented ASR feature set includes GPU-accelerated feature extraction, a beam search decoder over n-gram language models, CTC-based and neural voice activity detection, automatic punctuation, top-N transcripts from the beam decoder, word-level timestamps and confidences, inverse text normalization, offline and streaming speaker diarization, speech hints, Opus-encoded stream support, two-pass end-of-utterance detection and automatic speech translation.[5]

Two features are worth singling out. Streaming speaker diarization uses NVIDIA's Sortformer model running in parallel with ASR, tagging each word in a final transcript with a speaker ID; NVIDIA labels it a beta release supporting up to eight concurrent requests and available only with Parakeet-CTC and Conformer-CTC models in streaming mode.[3][5] Two-pass end of utterance lets a client receive an early intermediate transcript, controlled by a `stop_history_eou` parameter, while a slower and more accurate final transcript follows; NVIDIA's experiments suggest 240 ms as the setting, and the feature exists specifically so downstream language models can start work sooner.[5]

Riva's ASR documentation is unusually candid about limitations. Because Riva uses CTC-based acoustic models that "do not learn alignment during training," word timestamps "can be inaccurate."[3] Enabling neural VAD costs measurable performance: NVIDIA reports "a degradation in those metrics of the order of 10% to 15%" for Silero VAD.[3] Canary models run through PyTorch rather than TensorRT and therefore "offer a low throughput compared to other similar models"; Canary, Whisper and Distil-Whisper all lack word boosting.[3]

### Speech synthesis

Riva TTS is a two-stage pipeline. FastPitch and RadTTS++ generate a mel-spectrogram which HiFi-GAN converts to audio, while the newer MagpieTTS Multilingual model generates tokens that an audio codec model turns into speech.[8] Both streaming and offline inference are supported; streaming returns audio chunks as they are generated, which "significantly reduc[es] the latency (as measured by time to first audio) for large requests."[8]

Magpie-TTS was added in release 2.19.0 with English, Spanish and French and multi-speaker male and female emotion subvoices; 2.24.0 extended it to Italian, German, Vietnamese and Mandarin on Jetson, and 2.26.0 added Hindi and Japanese.[3] The model has known constraints: NVIDIA warns that long input "might produce truncated output" and recommends breaking text into normalized sentences of 400 characters or fewer.[3]

NVIDIA's headline TTS performance claim, published in the overview documentation, is that Riva "uses nonautoregressive models to deliver 12x higher performance on NVIDIA A100 GPUs compared to Tacotron 2 and WaveGlow models on NVIDIA V100 GPUs," and that a custom brand voice can be built "with only 30 minutes of an actor's voice data."[1][13] Both are vendor figures comparing across two GPU generations, not independent benchmarks.

### Translation

Riva's NMT service translates text between language pairs using models trained in NeMo.[9] The documentation enumerates 72 supported pairs, which is 36 non-English languages paired with English in both directions, covering Simplified and Traditional Chinese, Russian, German, European and Latin American Spanish, French, Danish, Greek, Finnish, Hungarian, Italian, Lithuanian, Latvian, Dutch, Norwegian, Polish, European and Brazilian Portuguese, Romanian, Slovak, Swedish, Japanese, Hindi, Korean, Estonian, Slovenian, Bulgarian, Ukrainian, Croatian and Arabic.[9] NVIDIA's developer page describes NMT as covering "up to 32 languages."[11]

Beyond text-to-text, Riva chains ASR, NMT and TTS into speech-to-text (S2T) and speech-to-speech (S2S) translation pipelines, and release 2.19.0 added the ability to point an S2T or S2S deployment at remotely deployed ASR or TTS servers.[3][9] A separate automatic speech translation path uses Whisper or Canary models to translate audio directly to text in a target language without an intermediate NMT step.[5] Clients can supply a custom translation dictionary, and `<dnt>` tags mark spans that should not be translated.[3][9]

## Models Riva serves

Riva serves NVIDIA's own speech model families rather than arbitrary third-party checkpoints. The following are drawn from the Riva user guide and, where noted, the NGC catalog. The [Parakeet](https://aiwiki.ai/wiki/parakeet) family (FastConformer encoders with CTC, RNN-T or TDT decoders) is the workhorse for streaming and high-throughput English and multilingual ASR, [Canary](https://aiwiki.ai/wiki/canary) covers offline multilingual recognition and translation, and OpenAI's [Whisper](https://aiwiki.ai/wiki/whisper) and its distilled and Japanese variants are supported for offline multilingual work.[3][5]

| Service | Models |
| --- | --- |
| ASR, streaming | Parakeet-CTC 0.6B and 1.1B, Parakeet-RNNT 1.1B (including unified multilingual code-switch), Conformer-CTC and Conformer-CTC-XL, Nemotron-ASR English |
| ASR, offline | Canary-1B, Canary-0.6B-turbo, Whisper large v3, Whisper large v3 turbo, Distil-Whisper large v3, Kotoba-Whisper v2.2 (Japanese) |
| Speaker diarization | Sortformer (streaming), MarbleNet VAD plus TitaNet embedding extractor (offline) |
| Voice activity detection | Silero VAD (recommended), MarbleNet VAD |
| TTS | FastPitch plus HiFi-GAN, RadTTS++, T5 TTS, MagpieTTS Multilingual plus audio codec, zero-shot TTS |
| NMT | Megatron 1B any-to-any; Riva Translate 1.6B and Riva-Translate-4B-Instruct are published as NIM containers rather than in the SDK documentation |
| NLP | BERT-base and BERT-large punctuation and capitalization models |

Pretrained ASR acoustic models, n-gram language models, lexicons, punctuation models and inverse text normalization grammars are published per language on the NGC catalog and referenced from the ASR documentation.[5] Riva can host several ASR models simultaneously, up to GPU memory, and selects among them from the `RecognitionConfig` of each client request.[5]

In 2026 NVIDIA began branding these models "Nemotron Speech," described as "open, state-of-the-art, production-ready enterprise speech models from the NVIDIA Speech research team for ASR, TTS, Speaker Diarization and S2S."[26] The lineage is visible in the model cards: `nvidia/nemotron-speech-streaming-en-0.6b` is a 600-million-parameter cache-aware FastConformer-RNNT with a 24-layer encoder, released under the NVIDIA Open Model License for commercial and non-commercial use, and reports 2.32% [word error rate](https://aiwiki.ai/wiki/word_error_rate) on [LibriSpeech](https://aiwiki.ai/wiki/librispeech) test-clean, 4.84% on test-other and a 6.93% average across eight datasets at a 1.12-second chunk size.[22] That is the same architecture family as Parakeet, repositioned under the [Nemotron](https://aiwiki.ai/wiki/nemotron) brand.

## Languages

NVIDIA's developer page describes the Nemotron Speech models served through Riva as covering "~40 languages" overall.[11] The per-service picture is narrower and worth stating precisely:

- **ASR with dedicated per-language pipelines:** English (US and GB), Spanish (US and Spain), German, Hindi, Russian, French, Brazilian Portuguese, Korean, Japanese, Arabic, Italian, Mandarin, Dutch (Netherlands and Belgium), plus Spanish-English, Japanese-English and EMEA code-switch models.[5] NVIDIA's marketing summary lists Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Mandarin, Portuguese, Russian and Spanish.[11]
- **ASR with multilingual models:** the universal and universal-RNNT Parakeet models accept 25 language codes and support automatic language identification when the client passes `multi`. NVIDIA marks a recommended subset of 13 for the universal-RNNT model.[5]
- **ASR via Whisper:** 100 language codes, with translation into English supported from all of them.[5]
- **TTS:** English, German, Italian, Mandarin and Spanish per NVIDIA's developer page, with Magpie-TTS adding French, Vietnamese, Hindi and Japanese across the 2.19.0 to 2.26.0 releases.[3][11]
- **NMT:** 36 non-English languages paired with English in both directions.[9]

Coverage is uneven in quality as well as breadth, and the documentation says so. Dutch models are "of beta quality," the Arabic acoustic model targets Modern Standard Arabic so "the accuracy for the Lebanese accent might be poor," Spanish (es-ES) and Italian Conformer-CTC-L models "have low throughput and high latency compared to other languages," and the Japanese-English code-switch model returns character rather than word timestamps.[3]

## Packaging as NIM microservices

Riva's data center distribution is now [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim). NVIDIA's speech NIM documentation states that "each NIM microservice packages a Nemotron model, the full NVIDIA inference stack (CUDA, TensorRT, Triton), and a unified API into a single container," exposing gRPC on port 50051 and HTTP on port 9000, with the model chosen at deploy time through `CONTAINER_ID` and `NIM_TAGS_SELECTOR` environment variables.[20]

The NGC catalog carries both service-level and model-level speech NIMs, including `nim/nvidia/riva-asr`, `nim/nvidia/riva-tts` and `nim/nvidia/riva-nmt`, alongside per-model containers such as `nim/nvidia/parakeet-1-1b-ctc-en-us`, `nim/nvidia/parakeet-0-6b-ctc-en-us`, `nim/nvidia/parakeet-tdt-0.6b-v2`, language-specific Parakeet builds for Spanish, Vietnamese, Mandarin and Taiwanese Mandarin, `nim/nvidia/nemotron-asr-streaming`, `nim/nvidia/magpie-tts-multilingual`, `nim/nvidia/fastpitch-hifigan-tts`, `nim/nvidia/megatron-1b-nmt` and `nim/nvidia/riva-translate-4b-instruct-v1.1`.[21]

This is the reverse of the original arrangement, in which the customer ran `riva-build` and `riva-deploy` to produce their own model repository. Riva release 2.24.0 recorded the change as a breaking one: "Deprecated x86 data center deployments. For x86 deployments, refer to Riva ASR NIM, Riva TTS NIM, and Riva NMT NIM documentation."[3] The Riva SDK path remains for embedded work on Jetson Thor, where NVIDIA still ships the build and deploy tooling.[4]

For evaluation without any local deployment, NVIDIA hosts the speech models behind managed endpoints on build.nvidia.com.[5][8][11]

## Licensing and how it is obtained

Riva is not open source. The server container is distributed from the NGC catalog under the NVIDIA Riva license, and the NGC listing states that "by pulling and using Riva software, you accept the terms and conditions of this license."[21][24] The documentation's licensing page now points at the NVIDIA AI Enterprise end user license agreement.[27] Production use is licensed through [NVIDIA AI Enterprise](https://aiwiki.ai/wiki/nvidia_ai_enterprise), and NVIDIA offers a 90-day free trial license as the documented route to free access.[11][25]

What is open is the periphery and the models:

- **Client libraries.** The `nvidia-riva` GitHub organization publishes the Python clients, C++ command-line clients, the protocol buffer definitions, the `nemo2riva` conversion tool and a WebSocket bridge, all under the MIT license, along with runnable tutorials and sample apps.[23]
- **Models.** The Parakeet checkpoints published between December 2023 and August 2025 are on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under CC-BY-4.0. Checkpoints published since then carry the NVIDIA Open Model License, which permits commercial use, and `nemotron-3.5-asr-streaming-0.6b` uses OpenMDW-1.1. The split follows the publication date rather than the Parakeet or Nemotron Speech name: five Parakeet-named repositories also carry the NVIDIA Open Model License.[22][26]
- **Training toolkit.** [NeMo](https://aiwiki.ai/wiki/nvidia_nemo), which trains and fine-tunes everything Riva serves, is an open-source PyTorch toolkit.[10]

The practical consequence is that a team can obtain the weights and the training code freely, and can call Riva through open client code, but the optimized serving path and its enterprise support are the licensed product.

## Performance

NVIDIA publishes latency and throughput measurements for Riva ASR taken with the `riva_streaming_asr_client` from the open `cpp-clients` repository, using a LibriSpeech dev-clean sample and the `--simulate_realtime` flag to imitate microphone input. Throughput is reported as RTFX, the ratio of audio duration to computation time. The on-premises numbers were collected with Riva 2.15.0 and the cloud numbers with 2.4.0.[7]

The default published configuration, a Conformer-CTC English (US) streaming low-latency pipeline on an [A100](https://aiwiki.ai/wiki/nvidia_a100) with 160 ms chunks, gives a sense of the shape of the curve:[7]

| Decoder | Concurrent streams | Average latency (ms) | p99 latency (ms) | Throughput (RTFX) |
| --- | --- | --- | --- | --- |
| n-gram beam | 1 | 13 | 40 | 0.999 |
| n-gram beam | 32 | 34 | 110 | 31.86 |
| n-gram beam | 128 | 86 | 360 | 126.5 |
| greedy | 1 | 12 | 30 | 1 |
| greedy | 32 | 32 | 100 | 31.9 |
| greedy | 128 | 80 | 330 | 126.5 |

NVIDIA reports a maximum of 218 effective streams with the n-gram language model and 223 without it for that configuration.[7] The overview documentation adds two broader marketing claims: that Riva can "offer real-time services that run in 150 milliseconds (ms) compared to the 25 seconds required on CPU-only platforms," and that stack-wide optimization "delivered 12X the gain versus the previous generation."[1][2] Neither is accompanied by a reproducible methodology in the way the streaming tables are.

### Published word error rates

NVIDIA publishes accuracy figures for Riva's ASR models, measured as word error rate for word-based languages and character error rate for character-based ones. The following are the offline figures on the datasets NVIDIA chose; lower is better.[7]

| Model | Language | Dataset | Best-latency WER (%) | Best-throughput WER (%) | Offline WER (%) |
| --- | --- | --- | --- | --- | --- |
| Parakeet-RNNT-1.1B | en-US | Common Voice 7.1 test | 10.74 | 10.54 | 9.77 |
| Parakeet-CTC-1.1B | en-US | Common Voice 7.1 test | 10.45 | 8.80 | 7.96 |
| Parakeet-CTC-0.6B | en-US | Common Voice 7.1 test | 10.57 | 8.87 | 8.45 |
| Canary-1B | en-US | Common Voice 7.1 test | not supported | not supported | 6.78 |
| Canary-0.6B | en-US | Common Voice 7.1 test | not supported | not supported | 8.65 |
| Parakeet-RNNT-1.1B | de-DE | MLS test | 11.30 | 9.16 | 7.09 |
| Parakeet-RNNT-1.1B | fr-FR | MLS test | 11.41 | 9.11 | 6.36 |
| Parakeet-RNNT-1.1B | ru-RU | RuLS test | 21.45 | 19.24 | 17.39 |
| Canary-1B | es-US | MLS test | not supported | not supported | 3.54 |

These are NVIDIA's own measurements on its own model deployments, and the streaming-configuration numbers are meaningfully worse than the offline ones, which is the expected trade. They are not directly comparable to the Open ASR Leaderboard figures often quoted for Parakeet, which use a different dataset suite.

## Customization

Riva's differentiator against hosted transcription APIs is that every stage of the pipeline can be modified. The documentation ranks seven techniques by effort:[6]

| Technique | Difficulty | What it does |
| --- | --- | --- |
| Word boosting | Quick and easy | Extends the vocabulary and raises the decoding score of a caller-supplied word list, at request time |
| Custom vocabulary | Easy | Extends the decoder's lexicon permanently to cover new words |
| Custom pronunciation (lexicon mapping) | Easy | Maps specific token sequences to specific words when a word has multiple pronunciations |
| Retrain the language model | Moderate | Trains a new n-gram language model on domain text |
| New inverse text normalization rules | Moderately hard | Changes how spoken forms are rendered in writing |
| Fine-tune an existing acoustic model | Moderately hard | Adapts the acoustic model with 10 to 100 hours of transcribed domain audio |
| Train a new acoustic model | Hard | Trains from scratch or with cross-language transfer, for a new language or dialect |

Word boosting is the mechanism most often reached for. A client passes a list of terms and a boost weight with the recognition request, and Riva raises those terms' scores during decoding. NVIDIA notes there is no limit on the number of boosted words and minimal latency impact even for tens of them, that out-of-vocabulary boosting works, and that boosting phrases or word combinations "is not yet fully supported (but do work)."[6] The documentation warns that boosting "can increase false positives" and recommends starting at a weight of 20 and increasing toward 100 while watching both word error rate and a terms error rate on the words of interest.[6] Custom vocabulary works differently: Riva's default Flashlight decoder is lexicon-based and "only emits words that are present in the provided vocabulary file," so a domain word absent from the lexicon has no chance of being produced at all until the vocabulary is extended.[6]

Fine-tuning goes back through NeMo. NVIDIA recommends fine-tuning ASR acoustic models "only with sufficient data approximately on the order of several hundred hours of speech," and suggests adapting the language model on in-domain text instead when that much audio is unavailable.[10]

## Deployments

### NVIDIA-published customer stories

NVIDIA has published several customer stories naming Riva. All are vendor-authored marketing with self-reported figures.

**T-Mobile** built Expert Assist, a call center application that transcribes customer conversations in real time and recommends solutions to frontline agents, initially for 17,000 customer service agents.[15][18] NVIDIA's case study attributes a 10 percent reduction in word error rate to fine-tuning with NeMo and a 10x latency reduction to Riva, "enabling the highest level of real-time performance for thousands of concurrent users."[18] The two contributions are attributed to different products, which is a distinction the summary headline of "10 percent higher accuracy, 10X overall latency improvement" flattens.

**RingCentral** fine-tuned NVIDIA speech models with NeMo and deployed them with Riva for real-time meeting transcription, reporting a word error rate reduction of over 10 percent and improved downstream summarization and sentiment analysis. Prashant Kukde, an associate vice president at RingCentral, is quoted saying the company can "transcribe meeting audio in real time with high accuracy while concurrently running thousands of streams."[19]

NVIDIA's September 2022 announcement additionally named Hewlett Packard Enterprise (ProLiant servers bundled with Riva), Deloitte (restaurant order-taking and healthcare chatbots), Interactions (its Curo platform, including a speech ordering system for 1-800-Flowers.com), Kore.ai (SmartAssist contact-center-as-a-service and its BankAssist, HealthAssist and AgentAssist products) and Quantiphi (closed captioning for media clients including Fox News).[15] An earlier NVIDIA post named Snap, T-Mobile, RingCentral and Kore.ai as Riva users in customer care, transcription and virtual assistants.[14]

### StudyFetch

In July 2026 NVIDIA published a case study titled "StudyFetch Cuts Inference Costs ~10x With NVIDIA, Expanding Access to AI Learning for Millions of Learners," and promoted it on X on July 31, 2026.[16][17] It is the most detailed public account of a Riva production deployment, and because it is NVIDIA marketing about NVIDIA's own customer, every figure in it is self-reported and none is independently measured.

StudyFetch is described as an AI-native learning workspace used by more than 7 million students, educators and lifelong learners, primarily in higher education and postgraduate programs, which turns live lectures, notes and course materials into AI tutoring, study plans, flashcards, practice tests and course-building curriculum.[16] Live lecture transcription is the foundation of the product: it is how the platform captures what a learner is taught.

NVIDIA says the company hit two constraints simultaneously while running live transcription on a major cloud provider's managed service at a volume of hundreds of thousands of college lectures per month. Cost was "already a six-figure monthly line item, and the unit economics were getting worse with every new learner." Accuracy was the second: real lectures are "full of domain-specific terminology in subjects like organic chemistry, anatomy, statistics, and programming," delivered by professors "with a wide range of accents and pacing," and the resulting transcript errors propagated into the AI tutor, the flashcards and the study guides.[16]

The rebuilt architecture, arrived at through the NVIDIA Inception startup program, has three parts.[16] Riva running the Parakeet ASR model produces live transcripts. NVIDIA NIM microservices package and deploy that pipeline across both AWS and Google Cloud, which NVIDIA says lets a small team "run production speech AI without a dedicated MLOps function." [Nemotron](https://aiwiki.ai/wiki/nemotron) open models sit alongside frontier models in StudyFetch's API layer, with [distillation](https://aiwiki.ai/wiki/knowledge_distillation) pipelines served through NIM to compress larger models into lower-latency deployables. The company's Learn Engine runs on [H100](https://aiwiki.ai/wiki/nvidia_h100) and [L40S](https://aiwiki.ai/wiki/nvidia_l40s) GPUs across more than 100 million learning interactions.

The reported result is precise about attribution: "Migrating live lecture transcription onto NVIDIA Riva and Parakeet running on NVIDIA L40S GPUs in NVIDIA NIM containers delivered roughly a 10x reduction in cost on the company's largest inference workload."[16] Ryan Trattner, StudyFetch cofounder and CTO, is quoted saying "StudyFetch exists to make great education accessible to anyone. NVIDIA is what makes that economically possible."[16]

Two qualifications matter for reading that number. First, it is a comparison between paying a third-party managed transcription API per unit and self-hosting GPU inference, not a like-for-like comparison of Riva against another speech engine on the same hardware. Anyone reading it as evidence that Riva is ten times cheaper than a competing ASR runtime is reading something the case study does not say. Second, the separate Nemotron distillation work is described as prospective rather than banked: NVIDIA writes that it "is on track to deliver similar substantial reductions across the rest of the inference stack," which is a forecast, not a result.[16] Likewise the Riva TTS pipelines for voice-based tutoring are "in development, targeted for summer rollout," not shipped.[16]

The case study also records that StudyFetch's new agentic learning platform, Honen, launched with NVIDIA and is reaching 250,000 K-12 students through AI literacy programs delivered alongside NVIDIA Training and Deep Learning Institute content, and that the company is evaluating three 8-way NVIDIA B300 systems to move conversational tutoring and course generation from closed models to open ones.[16]

## Positioning and limitations

Riva competes with hosted speech APIs from the large cloud providers and with self-hosted open-source stacks built around [Whisper](https://aiwiki.ai/wiki/whisper) and similar models. Its argument is control: the customer owns the deployment, the data does not leave their environment, and every layer from the lexicon to the acoustic model can be modified. NVIDIA's own framing emphasizes "data ownership and privacy" and deployment "on premises, in the cloud, at the edge, and in embedded devices."[2]

The counterweights are real. Riva only runs on NVIDIA GPUs and only serves models that have been through the Riva build path, so it is not a general-purpose speech server. The server is proprietary and, for production, gated behind an NVIDIA AI Enterprise license, even though the models it serves are openly licensed and could be run through NeMo or Hugging Face Transformers without it. Compilation binds a deployment to a target architecture. And NVIDIA has repeatedly retired platforms: Volta and earlier GPUs in 2.19.0, x86 data center deployment and Jetson Orin in 2.24.0, along with NVIDIA Fleet Command support.[3] A team standing up Riva on x86 today is not deploying the SDK at all; it is deploying speech NIM containers, which is a different operational model with a different upgrade cadence.

## See also

- [NVIDIA NeMo](https://aiwiki.ai/wiki/nvidia_nemo)
- [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim)
- [NVIDIA Triton Inference Server](https://aiwiki.ai/wiki/nvidia_triton_inference_server)
- [Parakeet](https://aiwiki.ai/wiki/parakeet)
- [NVIDIA Canary](https://aiwiki.ai/wiki/canary)
- [Speech recognition](https://aiwiki.ai/wiki/speech_recognition)
- [Automatic Speech Recognition Models](https://aiwiki.ai/wiki/automatic_speech_recognition_models)
- [Text-to-Speech](https://aiwiki.ai/wiki/text_to_speech_ai)
- [Neural Machine Translation](https://aiwiki.ai/wiki/neural_machine_translation)
- [AI Voice Agent](https://aiwiki.ai/wiki/ai_voice_agent)
- [NVIDIA AI Enterprise](https://aiwiki.ai/wiki/nvidia_ai_enterprise)
- [Edge AI](https://aiwiki.ai/wiki/edge_ai)

## References

1. NVIDIA. "Overview." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/overview.html
2. NVIDIA. "Riva." NVIDIA Riva user guide index. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/index.html
3. NVIDIA. "Release Notes." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/release-notes.html
4. NVIDIA. "Support Matrix." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/support-matrix/support-matrix.html
5. NVIDIA. "ASR Overview." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/asr/asr-overview.html
6. NVIDIA. "Basics of Speech Recognition and Customization of Riva ASR." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/asr/asr-customizing.html
7. NVIDIA. "Performance" (ASR). NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/asr/asr-performance.html
8. NVIDIA. "TTS Overview." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/tts/tts-overview.html
9. NVIDIA. "Translation Overview." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/translation/translation-overview.html
10. NVIDIA. "Overview" (custom model development and deployment). NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/model-overview.html
11. NVIDIA. "NVIDIA Nemotron Speech for Developers" (Riva developer page). https://developer.nvidia.com/riva
12. NVIDIA Developer Forums. "NVIDIA Jarvis has been renamed to NVIDIA Riva." July 28, 2021. https://forums.developer.nvidia.com/t/nvidia-jarvis-has-been-renamed-to-nvidia-riva/185025
13. NVIDIA. "NVIDIA Announces Riva Speech AI and Large Language Modeling Software for Enterprise." NVIDIA Technical Blog, November 9, 2021. https://developer.nvidia.com/blog/nvidia-announces-riva-speech-ai-and-large-language-modeling-software-for-enterprise/
14. NVIDIA. "Build Speech AI in Multiple Languages and Train Large Language Models with the Latest from Riva and NeMo Megatron." NVIDIA Technical Blog, March 28, 2022. https://developer.nvidia.com/blog/build-speech-ai-in-multiple-languages-and-train-large-language-models-with-the-latest-from-riva-and-nemo-megatron/
15. NVIDIA. "NVIDIA Riva Sets New Bar for Fully Customizable Speech AI." NVIDIA Blog, September 21, 2022. https://blogs.nvidia.com/blog/riva-speech-ai/
16. NVIDIA. "StudyFetch Cuts AI Inference Costs ~10X With NVIDIA." NVIDIA customer story. https://www.nvidia.com/en-us/case-studies/studyfetch/
17. NVIDIA (@nvidia). Post on X, July 31, 2026. https://x.com/nvidia/status/2083289093731934424 (text retrieved via https://api.fxtwitter.com/nvidia/status/2083289093731934424)
18. NVIDIA. "Speech AI for Award-Winning Customer Care" (T-Mobile). NVIDIA customer story. https://www.nvidia.com/en-eu/case-studies/speech-ai-for-award-winning-customer-care/
19. NVIDIA. "World-Class Speech AI for the Best Video Conferencing Experience" (RingCentral). NVIDIA customer story. https://www.nvidia.com/en-eu/case-studies/world-class-speech-ai-for-the-best-video-conferencing-experience/
20. NVIDIA. "NVIDIA Speech NIM Microservices Overview." https://docs.nvidia.com/nim/riva/asr/latest/overview.html
21. NVIDIA. "Riva Speech Skills." NGC catalog container listing. https://catalog.ngc.nvidia.com/orgs/nvidia/teams/riva/containers/riva-speech
22. NVIDIA. "nvidia/nemotron-speech-streaming-en-0.6b." Hugging Face model card. https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b
23. nvidia-riva. GitHub organization repositories. https://github.com/nvidia-riva
24. NVIDIA. "NVIDIA Riva License." https://developer.nvidia.com/riva/license
25. NVIDIA. "NVIDIA AI Enterprise Trial." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/ai-enterprise-trial.html
26. NVIDIA. "Nemotron Speech." Hugging Face collection. https://huggingface.co/collections/nvidia/nemotron-speech
27. NVIDIA. "End User License Agreement." NVIDIA Riva user guide. https://docs.nvidia.com/deeplearning/riva/user-guide/docs/reference/eula.html