TensorX Speech API

Basics

• Base URL: https://speech.inferix.ai
• Authentication: an API key from the API keys page, sent as Authorization: Bearer txs_.... WebSocket clients that cannot set headers can add ?api_key=txs_.... The Azure-style api-key header also works.
• Zero data retention: audio and text pass through memory only. We store timings, durations and status for usage, never content.
• POC limits: 20 calls at once per key (ask us to raise it), 90 minutes per call, 100 MB per file. Capacity is a shared NVIDIA B300.

Models

GET /v1/models lists models with lines in use and status.

Live: OpenAI Realtime-compatible

For clients that already use OpenAI or Azure OpenAI realtime transcription. Change the URL and the key; the events are the same.

wss://speech.inferix.ai/v1/realtime?model=nemotron-3.5-asr-streaming

• Send transcription_session.update (beta shape) or session.update (GA shape). Supported input formats: pcm16 24 kHz, g711_ulaw, g711_alaw, or GA audio/pcm with a rate, audio/pcmu, audio/pcma.
• Send audio with input_audio_buffer.append. The model finds sentence ends itself, so commit is optional.
• You receive conversation.item.input_audio_transcription.delta while people talk and ...completed per segment.
• OpenAI model names such as gpt-4o-transcribe map to our default model.

import asyncio, base64, json, websockets

async def main():
    url = "wss://speech.inferix.ai/v1/realtime?model=nemotron-3.5-asr-streaming"
    async with websockets.connect(url, additional_headers={"Authorization": "Bearer txs_..."}) as ws:
        await ws.send(json.dumps({"type": "transcription_session.update",
                                  "session": {"input_audio_format": "pcm16"}}))
        # send 24 kHz mono PCM16 chunks:
        # await ws.send(json.dumps({"type": "input_audio_buffer.append", "audio": base64.b64encode(chunk).decode()}))
        async for msg in ws:
            ev = json.loads(msg)
            if ev["type"].endswith("transcription.completed"):
                print(ev["transcript"])

asyncio.run(main())

Live: TensorX stream (simplest)

wss://speech.inferix.ai/v1/stream?model=nemotron-3.5-asr-streaming&sample_rate=16000&encoding=pcm_s16le&channels=1

• Send raw audio as binary frames, about 100 ms each. encoding: pcm_s16le, mulaw or alaw. sample_rate: 8000 to 48000.
• Two channels (agent and caller): set channels=2 and send interleaved stereo. Every event carries channel 0 or 1.
• Send {"type":"stop"} to finish. You get the last text, then done with usage.

{"type": "ready", "call_id": "call_...", "model": "...", "channels": 1}
{"type": "partial", "channel": 0, "text": "I can see there's an outstanding"}
{"type": "final",   "channel": 0, "text": "I can see there's an outstanding balance on your account."}
{"type": "done",    "call_id": "call_...", "usage": {"audio_seconds": 70.2, "first_text_ms": 640, "final_ms": 310}}
{"type": "error",   "code": "model_busy", "message": "..."}

Files

OpenAI-compatible upload. Most formats work: wav, mp3, m4a, ogg/opus, flac.

curl https://speech.inferix.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer txs_..." \
  -F model=nemotron-3.5-asr-streaming \
  -F file=@call.wav \
  -F response_format=json        # or text, verbose_json

Errors