Speech-to-Text (STT)

Speech-to-text (STT), also called automatic speech recognition (ASR), is AI technology that converts spoken language into written text. Every time you dictate a message, use a voice assistant, or see live captions on a video call, speech-to-text is at work.

How modern STT works

Early speech recognition used acoustic models that matched sound patterns to phonemes and assembled them into words. Modern STT systems use end-to-end deep learning — typically transformer-based models — that directly convert audio waveforms into text.

The process:

1. Audio is captured and converted into a spectrogram — a visual representation of sound frequencies over time.
2. The spectrogram is processed by a neural network that has been trained on thousands of hours of transcribed speech.
3. The model outputs a sequence of tokens that form the transcribed text.
4. Language model post-processing corrects likely errors and adds punctuation.

Key STT models

• Whisper (OpenAI): Open-source model with strong multilingual capabilities.
• Deepgram: Enterprise-focused with real-time transcription and customisation.
• Google Speech-to-Text: Cloud API with broad language support.
• AWS Transcribe: Amazon's managed transcription service.
• Azure Speech Service: Microsoft's offering with custom model training.

Factors affecting accuracy

• Audio quality: Clear audio in quiet environments produces the best results. Background noise, echo, and low-quality microphones degrade accuracy.
• Accents and dialects: Models perform better on accents well-represented in training data.
• Domain vocabulary: Technical, medical, or legal terminology may require custom models or vocabulary lists.
• Speaker overlap: Multiple speakers talking simultaneously challenges most systems.
• Language: High-resource languages (English, Mandarin, Spanish) have better recognition than low-resource languages.

Business applications

• Meeting transcription: Automatic notes and summaries from meetings and calls.
• Customer service: Transcribing and analysing support calls for quality and insights.
• Accessibility: Real-time captions for deaf or hard-of-hearing individuals.
• Content creation: Dictation for articles, emails, and documentation.
• Search: Making audio and video content searchable by transcribing it.
• Compliance: Creating audit trails of verbal communications in regulated industries.