NVIDIA Parakeet STT Now Available for Telnyx Voice AI

NVIDIA Parakeet is now available as a speech-to-text engine for Telnyx Voice AI, bringing self-hosted multilingual transcription with automatic language detection to production voice workflows. Built on the nvidia/parakeet-tdt-0.6b-v3 model, Parakeet transcribes 25 European languages on Telnyx infrastructure, with no third-party API hop and no language hint required.

What's new

• Multilingual STT engine: Parakeet transcribes speech across 25 European languages, including English, Spanish, French, German, Italian, Dutch, Polish, Portuguese, Russian, Ukrainian, and more, with automatic language detection. No language parameter is required.
• Self-hosted on Telnyx infrastructure: Parakeet runs on-network with no external API dependency, keeping transcription traffic inside Telnyx.
• 600M-parameter FastConformer-TDT model: Parakeet uses a FastConformer encoder with a TDT (Token-and-Duration Transducer) decoder, the same family that achieved strong English benchmarks on LibriSpeech (1.93% WER clean) and GigaSpeech (9.59% WER).
• Available on Voice API, TeXML, and WebSocket: Use Parakeet on Call Control transcription_start, TeXML <Gather> with transcription, and the standalone TeXML <Transcription> verb. REST batch transcription is not available.
• Final transcripts only, no endpointing: Parakeet returns final transcripts only, no interim or partial results. Endpointing is ignored. Accepts linear16, linear32, mp3, or wav audio at 16 kHz.

Why it matters

• Teams with mixed-language call traffic no longer need to pre-label the language on each call or maintain separate STT pipelines per language. Parakeet auto-detects.
• Self-hosted on Telnyx infrastructure, so the path between the caller and the transcription model never leaves Telnyx. That matters for regulated workloads where routing audio to a third-party STT vendor is a non-starter: HIPAA-aligned healthcare IVRs, EU data residency workloads, financial services call recording.
• Adds a 25-language option alongside the existing STT lineup (Deepgram, Whisper, Google, Azure, xAI, AssemblyAI, Speechmatics, Soniox) without changing the integration surface. Same Call Control, same WebSocket, same TeXML config. No integration tax.
• On-network processing: Parakeet runs inside Telnyx-hosted infrastructure, keeping audio and inference on the same private backbone. Each vendor boundary a call crosses adds 30 to 80ms. Parakeet removes one.
• Because Parakeet returns final transcripts only and ignores endpointing, it is not suited for real-time voice agent turn-taking. Use Deepgram flux or Soniox for end-of-turn detection. Parakeet is built for post-call transcription, audit pipelines, and mixed-language call recording.

Example use cases

• Customer support lines that receive calls in mixed European languages and need accurate transcription without per-call language tagging.
• Healthcare or financial IVRs where audio cannot be routed to a third-party STT vendor. Parakeet stays inside Telnyx infrastructure.
• Call recording and audit pipelines that need final transcripts for post-call analysis without streaming partials.
• Multilingual contact centers that need to log and search calls across European markets from a single STT pipeline.

Getting started

1. In Mission Control, go to AI > Assistants and open the assistant you want to configure.
2. Select the Voice tab and scroll to Transcription.
3. Set the transcription engine to Parakeet and the model to nvidia/parakeet-v3.
4. Save the assistant and test with mixed-language audio before routing production traffic.

For Call Control, POST to transcription_start:

{
  "transcription_engine": "Parakeet",
  "transcription_engine_config": {
    "transcription_engine": "Parakeet",
    "transcription_model": "nvidia/parakeet-v3"
  }
}

For TeXML <Gather>:

<Gather transcriptionEngine="Parakeet" model="nvidia/parakeet-v3">
  <Say>Please speak after the tone.</Say>
</Gather>

For WebSocket streaming:

wss://api.telnyx.com/v2/speech-to-text/transcription?transcription_engine=Parakeet&model=nvidia/parakeet-v3&input_format=linear16&sample_rate=16000

Parakeet accepts linear16, linear32, mp3, or wav audio at 16 kHz and returns final transcripts only. No interim or partial results. Endpointing is ignored.

Learn more in the Speech-to-Text models docs or the WebSocket streaming reference.