Speechmatics

The SpeechmaticsSTTService provides enterprise-grade, real-time speech-to-text using Speechmatics' Autonomous Speech Recognition (ASR). It features high accuracy, low latency, and advanced speaker diarization (who spoke when).

Installation

To use Speechmatics, install the required dependencies:

pip install "piopiy-ai[speechmatics]"

Prerequisites

• A Speechmatics account and API key (Get yours here).
• Set your API key in your environment:

  export SPEECHMATICS_API_KEY="your_api_key_here"

Configuration

SpeechmaticsSTTService Parameters

Parameter	Type	Default	Description
api_key	str	None	Speechmatics API key (defaults to env var).
base_url	str	None	WebSocket URL for Speechmatics ASR.
sample_rate	int	None	Audio sample rate in Hz.
params	InputParams	InputParams()	Advanced recognition and diarization settings.

InputParams

Parameter	Type	Default	Description
language	Language	EN	Language code for transcription.
operating_point	str	"enhanced"	Accuracy vs latency tradeoff (standard, enhanced).
enable_diarization	bool	False	Enable speaker identification.
turn_detection_mode	str	"fixed"	Endpoint handling (fixed, adaptive, smart_turn).
max_delay	float	None	Maximum delay for transcription results.

Usage

Basic Setup

import os
from piopiy.services.speechmatics.stt import SpeechmaticsSTTService
from piopiy.transcriptions.language import Language

stt = SpeechmaticsSTTService(
    api_key=os.getenv("SPEECHMATICS_API_KEY"),
    params=SpeechmaticsSTTService.InputParams(
        language=Language.EN,
        operating_point="enhanced"
    )
)

With Smart Turn Detection

from piopiy.services.speechmatics.stt import SpeechmaticsSTTService, TurnDetectionMode

stt = SpeechmaticsSTTService(
    api_key=os.getenv("SPEECHMATICS_API_KEY"),
    params=SpeechmaticsSTTService.InputParams(
        turn_detection_mode=TurnDetectionMode.SMART_TURN
    )
)

Notes

• Diarization: When enable_diarization is true, transcripts will include speaker labels (e.g., S1: Hello).
• Endpointing: SMART_TURN uses advanced ML models to detect when a user has finished speaking, providing a more natural interaction flow.