Soniox Speech-to-Text API
Product · API / developer platform by Soniox
Soniox Speech-to-Text (STT) API is a high-accuracy transcription service designed for real-time and asynchronous audio processing across more than 60 languages. It serves developers and businesses building voice agents, dictation tools, and meeting transcription platforms that require low-latency performance and native-speaker precision. The API features advanced capabilities such as speaker diarization for up to 15 speakers, automatic language identification, and semantic endpoint detection to improve turn-taking in conversational AI. Delivered via REST and WebSocket interfaces, it allows for
- Languages
- English, Spanish, French, German, Japanese, Chinese, Korean, Arabic, Afrikaans, Albanian, Hindi, Portuguese
- Supported Languages
- 60+
- Max Speakers (Diarization)
- 15
- Median Latency
- 249ms
- Word Error Rate (WER)
- 1.25%
- API Protocols
- REST, WebSocket
- Audio Token Conversion
- 1 hour ≈ 30,000 tokens
Overview
Soniox Speech-to-Text (STT) API is a high-accuracy transcription service designed for real-time and asynchronous audio processing across more than 60 languages. It serves developers and businesses building voice agents, dictation tools, and meeting transcription platforms that require low-latency performance and native-speaker precision. The API features advanced capabilities such as speaker diarization for up to 15 speakers, automatic language identification, and semantic endpoint detection to improve turn-taking in conversational AI. Delivered via REST and WebSocket interfaces, it allows for regional data residency in the US, EU, Japan, and India. Soniox STT distinguishes itself by handling complex audio environments, including code-switching and noisy backgrounds, at a significantly lower cost than traditional providers.
Core Capabilities
- Asynchronous transcription: Process recorded audio files asynchronously with high accuracy and speaker separation.
- Speaker diarization: Identify and label up to 15 different speakers in a single session for both live streams and recordings.
- Language identification: Automatically identify the spoken language at the token level and follow code-switching within a sentence.
- Context customization: Improve recognition of names, domain terms, and specialized vocabulary by providing context at request time.
- Semantic endpoint detection: Identify when a user has finished speaking using semantic cues rather than just silence.
Voice Agent Features
- Agent voice echo cancellation: Filter the agent's own voice out of the transcript to prevent false interruptions during conversations.
- Agent text input: Send agent replies to the model to improve transcription accuracy of short or ambiguous user answers.
- Framework integrations: Integrate natively with LiveKit, Pipecat, Twilio, LangChain, and other real-time voice frameworks.
- Regional deployment: Deploy in specific regions including the US, EU, Japan, and India to meet data residency requirements.
Pricing
- Async (file) transcription
- $1.50 per 1M tokens (approx. $0.10/hour)
- Real-time (streaming) transcription
- $2.00 per 1M tokens (approx. $0.12/hour)
- Input text tokens (context)
- $3.50 - $4.00 per 1M tokens
- Output text tokens
- $3.50 - $4.00 per 1M tokens
Specifications
- Supported Languages
- 60+
- Max Speakers (Diarization)
- 15
- Median Latency
- 249ms
- Word Error Rate (WER)
- 1.25%
- API Protocols
- REST, WebSocket
- Audio Token Conversion
- 1 hour ≈ 30,000 tokens
Added by AntSpace from public sources (docs.soniox.com).