Real-time voice with Gradium TTS & STT in TEN

We just shipped two new voice integrations for TEN: Gradium speech-to-text (STT/ASR) and Gradium text-to-speech (TTS). They unlock low-latency, ultrarealistic audio for assistants, copilots, and any product that needs to sound human while responding in real time. Explore Gradium at gradium.ai.

TL;DR

• Gradium STT streams transcriptions over WebSockets with interim and final results, plus VAD support to keep latency low.
• Gradium TTS generates lifelike PCM audio (48 kHz by default) and streams it back instantly for smooth turn-taking.
• Both ship as TEN extensions (gradium_asr_python and gradium_tts_python) with simple JSON config and region-aware endpoints (US/EU).
• Designed for real-time loops: microphone -> Gradium STT -> reasoning -> Gradium TTS -> speaker, all under one TEN app graph.

Why Gradium voice inside TEN

TEN already handles streaming I/O, session state, and routing between extensions. Gradium fits perfectly into that pipeline: its WebSocket APIs keep round trips tight, while the voice quality feels natural enough for production-grade customer experiences. Combining the two lets you prototype and ship responsive voice agents without gluing together separate services.

What we shipped

Gradium STT (ASR)

• WebSocket streaming for real-time transcription with interim and final results.
• Multi-region endpoints (wss://us.api.gradium.ai/api/speech/asr, wss://eu.api.gradium.ai/api/speech/asr).
• VAD-aware flow so you can trim silence and speed up responses.
• Flexible inputs: PCM, WAV, or Opus with 16-bit mono audio (24 kHz recommended).

Gradium TTS

• Streaming TTS that emits 16-bit PCM audio at 48 kHz by default, with options for 16 kHz and 24 kHz PCM.
• Works with any Gradium voice ID and model name; keep the style consistent across product surfaces.
• Region-aware WebSocket endpoint (wss://<region>.api.gradium.ai/api/speech/tts) plus simple GRADIUM_API_KEY auth.

Quick start: wire both into a TEN app

1. Set your API key:

export GRADIUM_API_KEY=your_gradium_key

2. Add Gradium STT and TTS nodes to your TEN graph (simplified example):

{
  "nodes": [
    {
      "type": "extension",
      "name": "gradium_asr",
      "addon": "gradium_asr_python",
      "extension_group": "gradium_asr_group",
      "property": {
        "params": {
          "api_key": "${env:GRADIUM_API_KEY|}",
          "region": "us",
          "model_name": "default",
          "input_format": "pcm",
          "sample_rate": 24000
        }
      }
    },
    {
      "type": "extension",
      "name": "gradium_tts",
      "addon": "gradium_tts_python",
      "extension_group": "gradium_tts_group",
      "property": {
        "params": {
          "api_key": "${env:GRADIUM_API_KEY}",
          "region": "us",
          "model_name": "default",
          "voice_id": "YOUR_GRADIUM_VOICE_ID",
          "output_format": "pcm"
        }
      }
    }
  ],
  "connections": [
    {
      "extension_group": "microphone_group",
      "extension": "microphone",
      "audio_frame_out": [
        {
          "name": "pcm_frame",
          "dest": [
            {
              "extension_group": "gradium_asr_group",
              "extension": "gradium_asr"
            }
          ]
        }
      ]
    },
    {
      "extension_group": "llm_router_group",
      "extension": "llm_router",
      "text_out": [
        {
          "name": "reply_text",
          "dest": [
            {
              "extension_group": "gradium_tts_group",
              "extension": "gradium_tts"
            }
          ]
        }
      ]
    }
  ]
}

3. Start your TEN app and you have a full duplex loop: mic -> Gradium STT -> LLM/tooling -> Gradium TTS -> speakers.

Optimization tips for low latency

• Match sample rates: keep STT input at 24 kHz mono PCM and TTS output at 48 kHz PCM for fidelity; downsample only once at the edge.
• Use VAD: let Gradium STT handle silence trimming to shorten end-of-utterance delays.
• Cache voices: pick a single Gradium voice_id per session to avoid extra lookups.
• Stay regional: choose the closest Gradium region (US/EU) for lower round-trip time.
• Stream small chunks: send ~80 ms PCM chunks (1,920 samples at 24 kHz) to keep transcripts flowing smoothly.

Where to use it

• Customer support and sales agents that need natural prosody and tight response loops.
• Real-time copilots in productivity or creative tools where hands-free operation matters.
• Multilingual kiosks, IVRs, and embedded devices that rely on reliable ASR plus high-quality playback.

Ready to try?

Gradium voice is live in TEN today. Grab an API key from gradium.ai, drop your config into property.json, and ship a voice experience that sounds human and responds fast. If you build something cool with Gradium TTS or STT, let us know; we'd love to feature it next.