What is Speech-to-Text (STT)? — AI Glossary

Speech-to-text (STT) — also called automatic speech recognition (ASR) — converts spoken audio into written text. It’s the technology behind voice assistants, live captions, meeting transcription tools, and voice commands. Modern AI-powered STT has reached human-level accuracy for many languages and accents, with tools like OpenAI’s Whisper making high-quality transcription available to anyone for free. Used alongside text-to-speech (TTS), STT enables fully voice-based AI interfaces.

How Modern STT Works

Early speech recognition used Hidden Markov Models (HMMs) and required speaker training — you had to “train” the software on your specific voice. Modern neural STT is speaker-independent and dramatically more robust:

• Feature extraction: Audio is converted to mel spectrograms (visual frequency-time representations).
• Encoder: A transformer encoder (self-attention-based) processes the spectrogram and creates rich audio representations.
• Decoder: A language model decoder generates the text transcript, token by token, using both the audio encoding and the previously generated text.

OpenAI’s Whisper (released 2022, open-source) is trained on 680,000 hours of multilingual audio from the internet. This breadth makes it remarkably robust to accents, background noise, technical vocabulary, and multiple languages — a step-change over earlier systems.

Leading STT Tools

The STT landscape spans free open-source tools to enterprise-grade APIs:

• OpenAI Whisper: Open-source, runs locally, excellent multilingual performance. Available via API at $0.006/minute.
• Deepgram: Real-time STT API with low latency, speaker diarization (identifying who spoke), and custom vocabulary.
• AssemblyAI: Strong accuracy, built-in summarization and sentiment analysis, speaker labels.
• Google Speech-to-Text: Strong multilingual support, tight Google ecosystem integration.
• Krisp: Combines noise cancellation with transcription, excellent for meeting environments (try Krisp).
• Wispr Flow: AI dictation that transcribes into any app using voice — productivity-focused (try Wispr Flow).

STT Applications Across Industries

STT is embedded in more workflows than most people realize:

• Productivity: Dictation apps, meeting transcription (Otter.ai, Fireflies), and voice-to-email.
• Accessibility: Live captions for the deaf and hard-of-hearing on Zoom, YouTube, and TV.
• Healthcare: Ambient clinical documentation (doctors speak notes, AI transcribes and structures EHR entries).
• Customer service: Call center analytics — transcribing every customer call for quality assurance and sentiment analysis.
• Legal: Court reporting, deposition transcription, legal dictation.
• Media: Automatic subtitle generation for video content.

The combination of STT + LLM processing + TTS creates conversational AI voice interfaces — what powers voice assistants like Siri and Alexa, but now increasingly accessible through APIs for any developer to build with.

Key Takeaways

• Modern STT converts speech to text using transformer models trained on hundreds of thousands of hours of audio.
• OpenAI Whisper made high-quality multilingual STT free and open-source.
• Speaker diarization (who said what) is now a standard feature of enterprise STT platforms.
• STT + LLM + TTS = the full voice AI stack powering conversational AI applications.
• Major applications include meeting transcription, accessibility, healthcare documentation, and call center analytics.

Frequently Asked Questions

How accurate is modern speech-to-text?

Top systems like Whisper achieve word error rates (WER) of 3-5% on standard benchmarks — comparable to professional human transcription. Accuracy drops in noisy environments, with heavy accents, or for technical/domain-specific vocabulary. Custom language model adaptation can recover much of this accuracy.

What is speaker diarization?

Diarization identifies who is speaking at any given moment in an audio recording. Instead of just a transcript, you get labeled segments: “Speaker A: …,” “Speaker B: …” This is critical for meeting transcriptions with multiple participants.

Can STT work in real time?

Yes. Real-time STT typically has a 100-500ms latency — fast enough for live captions and voice commands. Batch transcription (for pre-recorded audio) is more accurate because the model can use future context. Real-time and batch modes have different accuracy/latency tradeoffs.

What’s the difference between STT and voice assistants?

STT is just one component of a voice assistant. A full voice assistant pipeline includes: STT (speech to text) → intent detection or LLM processing (understanding what was asked) → backend action (search, device control, etc.) → TTS (generating spoken response). STT is the “ears”; TTS is the “mouth.”

Is Whisper free to use?

The Whisper model weights are free and open-source under MIT license — you can run it locally for free. OpenAI also offers Whisper via API at $0.006 per minute, which is among the most affordable commercial STT pricing available.

---

Want to go deeper? Browse more terms in the AI Glossary or subscribe to our newsletter for daily AI concepts explained in plain English.

Reduce background noise on every call: Krisp combines AI noise cancellation with transcription — great for remote workers.

You May Also Like

• What Is Artificial Intelligence
• AI Glossary
• AI for Dummies
• How Does AI Work
• Essential AI Skills

---

Get free AI tips daily → Subscribe to Beginners in AI

Sources

This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.

Last reviewed: April 2026