음성-텍스트 변환(STT)이란?
음성-텍스트 변환 기술의 이해, 작동 원리, 역사, 현대 AI가 자동 전사를 어떻게 변화시켰는지에 대한 종합 가이드.
음성-텍스트 변환 기술 이해하기
Speech to text (STT), also known as automatic speech recognition (ASR), is the technology that converts spoken language into written text. It allows computers to "listen" to human speech and produce a text transcript of what was said. STT systems are the backbone of voice assistants, closed captioning, dictation software, meeting transcription tools, and countless other applications we use every day.
At its core, speech to text solves a deceptively difficult problem: human speech is continuous, varies wildly between speakers, is affected by accents, background noise, speaking speed, and context. Turning that messy analog signal into clean, accurate text requires sophisticated algorithms that have been refined over decades of research.
Modern STT systems achieve accuracy rates above 95% for clear audio in major languages, rivaling human transcriptionists in many scenarios. This guide explains how that is possible, traces the history of the technology, and covers the different approaches used today.
음성-텍스트 변환의 작동 원리
Every speech-to-text system, whether classical or modern, follows a general pipeline. Audio comes in, gets processed through several stages, and text comes out. The stages differ in implementation, but the conceptual flow is consistent.
1. 오디오 전처리
Raw audio is first converted into a numerical representation the system can work with. This typically involves sampling the waveform (usually at 16 kHz for speech), applying noise reduction or normalization, and then extracting features. The most common feature representation is the mel-frequency cepstral coefficient (MFCC) or mel spectrogram, which transforms the audio into a time-frequency representation that mirrors how the human ear perceives sound. Modern neural models like Whisper use log-mel spectrograms computed from 25ms windows with 10ms stride.
2. 음향 모델
The acoustic model is the component that maps audio features to linguistic units. In classical systems, these units are phonemes (the smallest sound units of a language). The acoustic model answers the question: "Given this chunk of audio, what sound is being spoken?" Older systems used Gaussian Mixture Models (GMMs) combined with Hidden Markov Models (HMMs) for this task. Modern systems use deep neural networks -- recurrent neural networks (RNNs), convolutional neural networks (CNNs), or transformer architectures -- that directly learn the mapping from spectrograms to characters, subword tokens, or words.
3. 언어 모델
The language model provides linguistic context. It encodes the probability of word sequences in a given language. For example, "I went to the store" is far more probable than "Eye went two the store," even though they sound identical. The language model helps the system choose the correct words when the acoustics are ambiguous. Classical systems used n-gram language models trained on large text corpora. Modern end-to-end systems often have an implicit language model built into the neural network itself, though some still use external language models for rescoring.
4. 디코더
The decoder combines the outputs of the acoustic model and language model to produce the final transcript. It searches through the space of possible transcriptions to find the most likely one. Classical decoders used Viterbi search or weighted finite-state transducers (WFSTs). Modern systems often use beam search decoding with the neural network's output probabilities, or CTC (Connectionist Temporal Classification) decoding that handles the alignment between audio frames and output tokens automatically.
음성-텍스트 변환의 간략한 역사
The quest to make machines understand speech has spanned over seven decades, evolving from simple digit recognizers to today's near-human-level transcription systems.
1950~1970년대: 초기
The first speech recognition system, "Audrey," was built by Bell Labs in 1952. It could recognize spoken digits from a single speaker with about 97% accuracy. In 1962, IBM demonstrated "Shoebox" at the World's Fair, which could understand 16 English words. These systems were template-based: they stored reference patterns of speech and matched incoming audio against them. They were extremely limited -- single speaker, small vocabulary, isolated words only.
1980~1990년대: 통계적 방법
The introduction of Hidden Markov Models (HMMs) in the 1980s was transformative. Rather than matching templates, HMMs modeled speech as a statistical process, handling the variability of natural speech far better. The DARPA-funded research programs drove rapid progress, and by the 1990s, commercial products began to appear. Dragon Dictate (1990) was the first consumer speech recognition product, and Dragon NaturallySpeaking (1997) offered continuous speech recognition -- no more pausing between words. IBM ViaVoice and Microsoft Speech followed. These systems required extensive training on a specific user's voice and worked best in quiet environments.
2000~2010년대: 딥러닝 혁명
The application of deep neural networks to speech recognition, pioneered by Geoffrey Hinton's group around 2009-2012, led to dramatic accuracy improvements. Google adopted deep learning for its voice search in 2012, and error rates dropped by over 25% overnight. Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) networks, became the standard. Baidu's Deep Speech (2014) showed that a simple end-to-end neural architecture could match complex traditional pipelines. CTC loss functions made it possible to train models without pre-aligned transcripts.
2020년대: Transformer와 기반 모델
The transformer architecture, originally developed for text, was adapted for speech with spectacular results. Models like wav2vec 2.0 (Meta, 2020) introduced self-supervised pre-training for speech, learning useful representations from unlabeled audio. OpenAI's Whisper (2022) was a watershed moment: trained on 680,000 hours of multilingual audio from the web, it delivered robust transcription across 100+ languages and noisy conditions without any fine-tuning. NVIDIA's Canary and Parakeet models pushed the boundaries further with CTC and transducer architectures optimized for production use. Today, the best models achieve word error rates under 5% on standard benchmarks, approaching human parity.
음성-텍스트 변환 활용 사례
STT 접근법 비교
Over the decades, three main approaches to speech recognition have emerged. Each represents a different generation of the technology.
| Approach | How It Works | Strengths | Weaknesses |
|---|---|---|---|
| Rule-Based / Template | Matches input audio against stored templates using dynamic time warping or hand-crafted rules. | Simple to implement; works well for tiny vocabularies (digits, commands). | Cannot scale to large vocabularies; no adaptation to new speakers or noise; effectively obsolete. |
| HMM / Statistical (GMM-HMM) | Models speech as a sequence of hidden states. GMMs model emission probabilities; HMMs model temporal transitions. Separate acoustic model, language model, and pronunciation dictionary. | Well-understood mathematical framework; modular (components can be improved independently); dominated from 1980s to 2012. | Requires expert feature engineering; limited ability to learn complex patterns; lower accuracy than neural approaches. |
| Neural / Transformer (End-to-End) | A single neural network (or encoder-decoder pair) maps audio directly to text. Architectures include CTC, RNN-Transducer, attention-based seq2seq, and transformer. Trained on massive datasets. | Highest accuracy; learns features automatically from data; handles noise and accents well; multilingual models possible; benefits from scale. | Requires large training data and compute; can be a black box; latency can be higher for large models; may hallucinate on silence. |
Today, virtually all production STT systems use neural approaches. The transformer architecture has become dominant, with models like Whisper (encoder-decoder with attention), Canary (CTC/transducer hybrid), and Parakeet (CTC with fast-conformer) leading the field. The choice between them often comes down to the trade-off between accuracy, latency, and computational cost.
STT.ai 작동 방식
STT.ai is a transcription platform that gives you access to multiple state-of-the-art speech recognition models through a single interface. Rather than locking you into one model, STT.ai lets you choose the best model for your specific needs.
1. 업로드 또는 녹음
Upload any audio or video file (MP3, WAV, MP4, MKV, and 20+ more formats), record directly from your microphone, or paste a URL from YouTube, Vimeo, or any platform. Files up to 500MB are supported.
2. 모델 선택
Select from 10+ AI models including Whisper Large v3, Whisper Turbo, Distil-Whisper, NVIDIA Canary, and Parakeet. Each model has different strengths -- accuracy, speed, language coverage, or specialized domain performance. Or let STT.ai auto-select the best one.
3. 트랜스크립트 받기
Transcription runs on GPU-accelerated servers and typically completes in seconds. The result includes word-level timestamps, speaker identification, and can be exported as TXT, SRT, VTT, DOCX, JSON, or PDF. Share with a link or download directly.
STT.ai supports 100+ languages with automatic language detection, provides speaker diarization (identifying who said what), and offers both a web interface and a REST API for developers. The platform includes a generous free tier — 10 free minutes a day with no signup, and 600 free minutes when you create an account.
핵심 지표: STT 정확도 측정 방법
The standard metric for evaluating speech-to-text systems is the Word Error Rate (WER). WER is calculated as:
WER = (Substitutions + Insertions + Deletions) / Total Words in Reference
A WER of 5% means that 5 out of every 100 words are incorrect. Human transcriptionists typically achieve 4-5% WER on conversational speech. The best AI models now achieve comparable or better performance on clean audio, though challenging conditions (heavy accents, background noise, multiple overlapping speakers) can increase error rates significantly.
Other metrics include Character Error Rate (CER), useful for languages without clear word boundaries like Chinese or Japanese, and Real-Time Factor (RTF), which measures how fast the system processes audio relative to the audio duration (RTF < 1 means faster than real-time).
음성-텍스트 변환의 미래
Speech to text technology continues to advance rapidly. Several trends are shaping its future:
- Multimodal models that combine audio, video, and text understanding are emerging, enabling lip-reading-assisted transcription and better handling of ambiguous speech.
- On-device processing is becoming more feasible as models are compressed and optimized. This enables private, offline transcription on phones and laptops without sending audio to the cloud.
- Low-resource languages are benefiting from self-supervised learning and multilingual transfer, bringing STT to languages that previously had too little training data.
- Real-time streaming with sub-second latency is improving, making live captioning and simultaneous translation more practical.
- Personalization through few-shot adaptation allows models to quickly learn a user's speaking style, vocabulary, and accent preferences.