𝗩𝗼𝗶𝗰𝗲 𝗔𝗜: 𝗧𝗧𝗦 - 𝗚𝗶𝘃𝗶𝗻𝗴 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗮 𝗩𝗼𝗶𝗰𝗲

Source: Dev.to

The Transformation

Input: “Great news! Your flight to Paris is confirmed.”

Output: (audio waveform)

The TTS Pipeline

1️⃣ Text Analysis

• “How to pronounce this?”
• Normalization ($50 → “fifty dollars”)
• Grapheme‑to‑phoneme conversion
• Homograph resolution (e.g., read vs read)

2️⃣ Prosody Prediction

• How should it sound?
• Pitch contour (intonation)
• Duration (speed)
• Stress & emphasis
• Pauses

3️⃣ Acoustic Model

• Generate mel spectrogram
• Models: Tacotron 2, FastSpeech 2, VITS
• Maps phonemes → audio features

4️⃣ Vocoder

• Convert to audio waveform
• Technologies: HiFi‑GAN, WaveGlow, WaveNet
• Spectrogram → actual audio

🎯 And that closes the loop:
Listen → Think → Speak

That’s the full Voice AI pipeline.