NVIDIA Riva

​

Overview

NVIDIA Riva provides two STT services:

• RivaSTTService for real-time streaming transcription using Parakeet models
• RivaSegmentedSTTService for segmented transcription using Canary models with advanced language support

​

Installation

To use NVIDIA Riva services, install the required dependency:

pip install "pipecat-ai[riva]"

You’ll also need to set up your NVIDIA API key as an environment variable: NVIDIA_API_KEY.

​

Frames

​

Input

• InputAudioRawFrame - Raw PCM audio data (16-bit, mono)
• STTUpdateSettingsFrame - Runtime transcription configuration updates
• STTMuteFrame - Mute audio input for transcription

​

Output

• InterimTranscriptionFrame - Real-time transcription updates (streaming only)
• TranscriptionFrame - Final transcription results
• ErrorFrame - Connection or processing errors

​

Service Comparison

​

Models

See NVIDIA’s model cards for detailed performance metrics.

​

Language Support

​

RivaSTTService (Parakeet)

Primarily supports English with various regional accents:

• Language.EN_US - English (US) - en-US

​

RivaSegmentedSTTService (Canary)

Supports multiple languages:

​

Usage Example

​

Real-time Streaming (RivaSTTService)

from pipecat.services.riva.stt import RivaSTTService
from pipecat.transcriptions.language import Language

# Basic streaming configuration
stt = RivaSTTService(
    api_key=os.getenv("NVIDIA_API_KEY"),
    params=RivaSTTService.InputParams(
        language=Language.EN_US
    )
)

# Use in pipeline for real-time conversation
pipeline = Pipeline([
    transport.input(),
    stt,
    context_aggregator.user(),
    llm,
    tts,
    transport.output(),
    context_aggregator.assistant()
])

​

Segmented Multi-language (RivaSegmentedSTTService)

from pipecat.services.riva.stt import RivaSegmentedSTTService
from pipecat.audio.vad.silero import SileroVADAnalyzer

# Multi-language segmented transcription
stt = RivaSegmentedSTTService(
    api_key=os.getenv("NVIDIA_API_KEY"),
    params=RivaSegmentedSTTService.InputParams(
        language=Language.ES,  # Spanish
        profanity_filter=False,
        automatic_punctuation=True,
        boosted_lm_words=["inteligencia", "artificial"],
        boosted_lm_score=5.0
    )
)



# Use in pipeline
pipeline = Pipeline([
    transport.input(),
    stt,  # Processes audio segments
    context_aggregator.user(),
    llm,
    tts,
    transport.output(),
    context_aggregator.assistant()
])

​

Advanced Configuration

Both services support advanced ASR parameters:

​

Word Boosting

• boosted_lm_words: List of domain-specific terms to emphasize
• boosted_lm_score: Boost intensity (default: 4.0, recommended: 4.0-8.0)

​

Audio Processing

• profanity_filter: Filter inappropriate content
• automatic_punctuation: Add punctuation automatically
• verbatim_transcripts: Control transcript formatting

​

Voice Activity Detection (Streaming only)

• start_history: History frames for speech start detection
• start_threshold: Confidence threshold for speech start
• stop_threshold: Confidence threshold for speech end

​

Metrics

• Time to First Byte (TTFB) - Latency from audio segment to transcription
• Processing Duration - Time spent processing each segment

​

Additional Notes

• Authentication: Uses NVIDIA Cloud Functions with Bearer token authentication
• Real-time vs Batch: Choose streaming for conversation, segmented for accuracy and multi-language
• VAD Requirement: Segmented service requires Voice Activity Detection in the pipeline
• Custom Endpoints: Supports custom Riva server endpoints for on-premise deployments