MurmuraText-to-speech sounds dramatically better than it did a decade ago because the underlying stack changed. Instead of stitching together recorded fragments or relying on hand-tuned signal processing, modern systems predict prosody and generate audio with neural networks trained on large speech corpora.[1][2]
This article is a technical overview for readers, product teams, and accessibility practitioners who want the plain-English version of what is happening under the hood in 2026.
Brief history: from concatenative to neural
Concatenative TTS: good recordings, rigid output
Older high-quality systems were often concatenative: they stored thousands of recorded speech fragments, then stitched them together at runtime. When the source recordings were excellent, the result could sound natural. The downside was flexibility. Changing emotion, emphasis, pacing, or speaker identity usually meant collecting and curating an entirely new voice database.[1]
Parametric TTS: more controllable, less natural
The next wave used statistical or parametric models. Those systems were easier to control and lighter to deploy, but many listeners remember them as the era of flattened intonation and metallic timbre. They were useful, especially for accessibility, but rarely convincing.
Neural TTS: spectrograms, vocoders, and much better prosody
The modern shift came when neural systems started predicting richer acoustic representations directly from text. Tacotron 2 is the canonical example: a text-to-mel-spectrogram model paired with a neural vocoder. That architecture showed how much naturalness could improve once the model learned rhythm, phrasing, and pronunciation jointly instead of through mostly hand-built rules.[2]
Later families such as FastSpeech and FastSpeech 2 made synthesis faster and more stable by moving away from strictly autoregressive generation. That mattered for real products because low latency is not a cosmetic feature in TTS; it is the difference between a voice assistant feeling instant or laggy, and a reading app feeling playable or broken.[3]
How a modern neural TTS pipeline works
1. Text normalization
The system first expands raw text into something speakable. Dates, abbreviations, currency, URLs, section numbers, and initials all need to be resolved. “Dr.”, “2026”, and “$19.99” are not audio yet; they are normalization problems.
2. Linguistic and prosody planning
Next, the model decides how the text should be spoken: where pauses belong, which words deserve emphasis, how a sentence should rise or fall, and how quickly each token should unfold. This stage is why two providers can read the same paragraph with very different levels of clarity or warmth.
3. Acoustic prediction
Most production systems predict an intermediate acoustic representation such as a mel spectrogram rather than jumping straight to waveform samples. That representation is compact, fast to model, and expressive enough to preserve timing, pitch, and timbre.[2]
4. Neural vocoding
A vocoder or decoder turns that acoustic representation into audible waveform. WaveNet popularized the idea that directly modeling raw audio with deep neural networks could close much of the quality gap between synthetic and recorded speech. Modern vocoders are usually far faster than the original WaveNet implementations, but the core idea remains influential.[1]
5. Streaming, chunking, and caching
Most user-facing apps do not wait for a whole chapter before they start playback. They synthesize and stream in chunks, cache completed audio, and keep the decoder busy ahead of the playhead. Good TTS products therefore depend on systems design as much as model quality.
What “edge-based” neural TTS means in practice
In 2026, “edge” usually means one of three things: the model runs fully on device, the heavy work runs on a nearby edge node with low round-trip latency, or the product uses a hybrid model that streams short chunks while caching locally. Apple’s Personal Voice is a clear on-device example: Apple documents that recorded speech is processed securely on device before the resulting voice can be shared across a user’s Apple devices.[5]
The reason edge delivery matters is simple: latency, privacy, and resilience. On-device or near-device synthesis reduces waiting, narrows the amount of data that has to travel over the network, and makes offline or low-connectivity experiences more realistic.
Why quality still varies between providers
Training data quality
A model trained on clean, expressive, well-aligned speech usually wins. A large voice library helps, but the recording conditions, labeling quality, and speaker consistency often matter more than raw hours alone.
Latency budget
Some providers optimize for instant playback. Others spend more compute per utterance to get richer expressiveness. That is why fiction, navigation prompts, short assistant replies, and long-form document reading can all favor different engines.
Language coverage and text normalization
The hard part is not only sounding human in English. It is handling mixed-language text, acronyms, tables, citations, code snippets, and edge-case punctuation without collapsing. A provider can have an excellent demo voice and still perform poorly on messy real-world documents.
Product choices above the model
Pronunciation lexicons, sentence segmentation, buffering strategy, speed controls, and highlighting sync all change how users perceive “voice quality.” Two apps can call the same backend and still feel very different because their playback and reading UX are different.
Where the stack is heading
The broad direction is clear: smaller models, faster inference, better multilingual coverage, and more hybrid delivery. The most useful products will not necessarily be the ones with the single most theatrical voice sample. They will be the ones that combine strong models with reliable document parsing, low latency, predictable pricing, and careful accessibility design.
Sources
- DeepMind: “WaveNet: A Generative Model for Raw Audio”
- Shen et al., “Tacotron 2: Generating Human-like Speech from Text”
- Ren et al., “FastSpeech 2: Fast and High-Quality End-to-End Text to Speech”
- Microsoft Learn: Azure Speech text-to-speech overview
- Apple Support: Create a Personal Voice on iPhone, iPad, or Mac
About the author
This explainer was published by the Murmura editorial team, which writes about document reading, accessibility, and text-to-speech product design.