Text-to-Speech Technology:
From Concatenative to Neural AI

The synthesis pipeline (what actually happens)

Speech synthesis turns written tokens into a waveform. A website player is only the last hop: play a file. The interesting work is earlier.

• Text in - isolated article, highlighted span (RHT), or a translated string. Isolation is product work, not a vocoder paper.
• Linguistic analysis - words, stress, pauses. Smart Pronunciation Aliases sit here: brand strings the model would otherwise guess.
• Acoustic model - concatenative units, a parametric vocoder, or a neural net that predicts audio.
• Waveform out - stored once under Generate Once, Keep and Play, then streamed. Browser speechSynthesis skips this store and uses whatever the OS shipped.

Three families of TTS models

Neural systems sit on top of that history. Rule-based and statistical stacks still exist in textbooks. What you hear on GSpeech commercial voices is AI synthesis: OpenAI, Gemini, Google Cloud Neural2 / Polyglot / Chirp3 HD, and WaveNet - not a local concatenative bank on your PHP host.

WaveNet: modelling the raw waveform

WaveNet changed the parametric story by predicting the audio signal sample by sample instead of handing a compact spectrogram to a vocoder. Google’s 2016 work showed that a dilated convolutional net can cover thousands of timesteps. Raw audio is dense - typically 16,000 samples per second or more - so autoregressive generation is expensive, but it captures lip smacks and speaker quirks concatenative glue often misses.

PixelRNN and PixelCNN showed that generating an image one channel at a time was possible. WaveNet is the 1D version of that idea: a fully convolutional network whose dilation factors grow the receptive field exponentially.

At training time the inputs are real recorded waveforms. At sampling time each step draws from the network’s predicted distribution, feeds that sample back, and continues. That loop is why early WaveNet was research-only compute; later stacks slimmed it into production Google TTS.

MOS and the 2016 gap

Google evaluated WaveNet against then-best parametric and concatenative systems using Mean Opinion Scores (MOS) - blind human ratings (over 500 ratings on 100 test sentences in the original report). WaveNet cut the remaining gap to human speech by more than half for US English and Mandarin Chinese in those tests. That is a research milestone, not a GSpeech marketing score.

WaveNet is one family in the Unified AI Voice Engine. Later stacks - Gemini TTS, Chirp3 HD, Neural2, Polyglot, and OpenAI - are what you pick as a voice in the Cloud Console. You do not deploy a WaveNet checkpoint on WordPress.

Gemini TTS: prompt-controlled speech

WaveNet predicted the next audio sample. Gemini TTS is a later Google speech-generation path: you describe how the read should sound in natural language - style, accent, pace, tone, emotion, delivery - and the model performs that direction. That is why GSpeech custom voice prompts and Voice Vibes (styles, not extra voice IDs) matter on Gemini: the control sits in the prompt, not in a new concatenative bank.

Linguistic cleanup still happens first. Smart Pronunciation Aliases can expand an acronym or lock a brand token; Gemini then speaks the prepared string with the requested delivery. Official API notes: Gemini TTS.

Gemini Pro

Gemini Pro is the ultra-realistic Gemini TTS model inside GSpeech. You write a custom voice prompt the way you would brief a narrator: calmer product explainer, brighter blog, tighter accent, slower pace, more emotion. The visible HTML stays clean for readers and SEO. There is no separate Gemini key, no Google AI Studio project, no extra provider console.

Technically this is still cloud synthesis, then a stored file. Repeat listens are playback, not a fresh Gemini round-trip. Use Pro when the audio is the experience - a hero article or a product page where delivery should feel directed, not generic.

Gemini Flash

Gemini Flash is the same Gemini TTS family, built for speed. Prompt vocabulary matches Pro (style, accent, pace, tone, emotion, delivery), so you can draft a voice direction once and reuse it while Flash turns pages around faster. Catalog workhorse: many posts, WooCommerce descriptions, multilingual drafts. Promote a hero URL to Pro when you want the last bit of performance.

Chirp3 HD: Google Cloud’s current HD neural family

Chirp3 HD is not “WaveNet with a new name.” WaveNet (2016) modelled raw audio with dilated convolutions. Neural2, Polyglot, and WaveNet voices still sit in the Google Cloud catalog (the 230+ floor: Standard, WaveNet, Neural2, Polyglot, Journey, Studio, News). Chirp3 HD is Google’s later high-definition Cloud TTS generation - built for realism and natural intonation rather than a flat read-aloud. Official notes: Chirp 3 HD.

Inside GSpeech, Chirp3 HD is a Cloud Console voice: no Google Cloud project, no service-account JSON. You pick the voice; synthesis runs off your origin; the file is stored and played through Full Page, Button, Circle, Floating, Context, or RHT. Speed and pitch stay in the console. Brand names still belong in Smart Pronunciation Aliases - that layer is independent of which Google family you chose.

When to pick Chirp3 HD vs Gemini: Chirp3 HD when you want production Google Cloud HD neural quality and a consistent device-independent read. Gemini Pro / Flash when you want natural-language prompt control over style and emotion. Both are commercial Unified AI Voice Engine paths. Engine total is 300+ voices across 98 languages (Gemini ~30 × 3 for Chirp 3 HD, Flash, and Pro per language, plus OpenAI 11 × 2, plus the Google Cloud catalog). Voice Vibes are styles, not extra IDs.

• Neural2 - widely used production neural voices in the same Google Cloud catalog.
• Polyglot - stronger multilingual / cross-language performance.
• WaveNet - the 2016 raw-waveform family, still available as catalog voices, not the only Google path.

Unified AI Voice Engine vs browser TTS

Device speechSynthesis is a local API. Voices differ by OS. There is no Generate Once, Keep and Play file, no Smart Audio Sync when the CMS copy changes, and no Voice Vibes prompt layer. GSpeech Cloud Processing runs the model in the cloud, stores the file, and the widget plays it. Temporary generation issues do not delete previously generated files. Files are not removed when an account returns to the Free plan. New generation still follows the active plan.

models:  "OpenAI, Gemini, Google Cloud, Chirp3 HD, Neural2, Polyglot, WaveNet"
langs:    "98 languages, 300+ voices (commercial)"
control: "Voice Vibes + Smart Pronunciation Aliases"
keys:    "No API Keys Required"

Voice Vibes (commercial) are prompt-side style, not a new concatenative bank. Aliases fix how a token is spoken without changing the visible HTML. Both are control layers on top of whatever acoustic model the engine selected for that voice.

From paper to a voice in the console

Concatenative banks, parametric vocoders, and WaveNet’s raw-waveform bet are the lineage. Gemini TTS (Pro / Flash) adds prompt-controlled delivery; Chirp3 HD is Google Cloud’s current HD neural family. On a GSpeech site you hear the current Unified AI Voice Engine through a widget - Full Page, Button, Circle, or Read Highlighted Text (RHT) - with files kept after first generation. For “should my URL have a player,” use the benefits article. For wp-admin clicks, use the how-to.