Speech-to-Text (STT)

Speech-to-text (STT), also called automatic speech recognition (ASR), converts spoken audio into written text. Modern STT systems (OpenAI Whisper, Google Speech-to-Text, AssemblyAI, Deepgram) use end-to-end neural architectures — typically transformer encoder-decoder models — trained on hundreds of thousands of hours of multilingual audio, achieving word error rates under 5% for clear studio audio in major languages and holding up reasonably well against accents, background noise, and overlapping speakers — a substantial jump in robustness compared to the pre-2020 generation of ASR systems that required near-studio-quality audio to perform reliably. Why it matters for SaaS builders: STT is the entry point for any voice-driven product — meeting transcription tools (Otter.ai, Fireflies), voice-controlled apps, call-center analytics, subtitle/caption generation, and voice-to-text input fields. It's also a critical preprocessing step for voice AI agents, where STT converts the caller's speech to text, an LLM reasons over that text, and TTS speaks the response back — the classic "STT → LLM → TTS" voice-agent pipeline. Integration is typically a single API call: upload an audio file or stream audio in real time and receive a transcript, often with word-level timestamps, speaker diarization (who said what), and confidence scores. A concrete worked example — building a "meeting notes" SaaS feature: (1) user uploads a 45-minute Zoom recording (MP3), or the app captures audio live via the Zoom/Meet bot API for real-time processing; (2) the app calls `POST https://api.deepgram.com/v1/listen?model=nova-2&diarize=true&punctuate=true` with the audio file; (3) the response returns a JSON transcript with per-speaker segments, word-level timestamps, and confidence scores, e.g. `{"speaker": 0, "start": 12.4, "confidence": 0.97, "text": "Let's review Q3 targets."}`; (4) the app feeds the full transcript into an LLM with the prompt "Summarize this meeting into action items grouped by owner, and flag any unresolved questions"; (5) the summary, action-item list, and full searchable transcript are shown side by side, with each action item deep-linking to the exact transcript timestamp it was extracted from so a user can verify context. Key parameters: language/model selection, diarization (speaker separation — critical for multi-person calls but adds processing time and cost), real-time streaming vs. batch processing, and custom vocabulary boosting to improve recognition accuracy on domain jargon like product names or technical acronyms that a general-purpose model would otherwise mishear. Cost is typically billed per audio-minute processed, so high-volume products often pre-filter silence using voice activity detection before sending audio to the paid STT API.