Streaming synthesis commits audio before all future text is known

A sentence-level model may rely on future text for pronunciation and prosody even if its neural computation is parallel. Cutting the input string into short pieces does not automatically make that model a context-correct streaming synthesizer. Incremental approaches such as the prefix-to-prefix work by Ma and colleagues explicitly trade text lookahead against early speech production. Their results support that studied design rather than a universal guarantee that any bounded-lookahead system matches an unrestricted offline voice. State which future words or tokens must arrive before a segment becomes stable.

Track time to stable input prefix, acoustic-model output, first vocoder samples, network delivery, and actual audible playback. Some stages overlap after startup while others are on the first-chunk critical path. A stream can have excellent average compute throughput yet stall when its next text prefix arrives late. Conversely, a fast first chunk can be followed by underruns if later work cannot keep pace. Measure both initial delay and inter-chunk continuity under realistic text arrival, not only synthesis of a complete sentence already available in memory.

An acoustic model can carry hidden or attention state, and a vocoder can need waveform or feature context around a chunk edge. Preserve the supported state and define whether future context is available before rendering. Generating independent waveform pieces and concatenating them can introduce discontinuities, inconsistent pitch, or repeated boundary samples. A justified overlap/crossfade rule needs aligned sample ownership and the model's synthesis convention. It cannot repair a wrong pronunciation or phrase interpretation that was already committed because insufficient future text was available.

Prosody is not a scalar side effect of chunk duration. Stress, phrase breaks, speaking rate, and intonation can depend on the larger utterance and dialog context. Evaluate ambiguous phrases, numbers, abbreviations, long sentences, and text that arrives with later punctuation changes. Define what the system does with revised unspoken text and how it rejects stale audio from an older revision. Avoid silently regenerating already played content or duplicating it as a repair. A stable voice identity and a smooth waveform join do not by themselves establish natural or semantically appropriate phrasing.

Report first audible output, underflow frequency, longest gap, total latency, and listening judgments on matched content. Include text arrival delays, cancellation, retry, and output-device buffering in the test. Compare at the same voice, language, speed, and quality settings; reducing the output duration can change a ratio without improving pronunciation or interaction. A perceptual proxy or automatic similarity score can guide development but does not replace appropriate listening assessment, and a favorable average should not hide problematic languages, devices, or long-form sequences.

The queue lab models only a smooth audio-supply imbalance. Reset, explicitly set BOTH production λ=1.2 and consumption μ=1, predict q(2)=4.4 from initial backlog four, then Play or Step four times. Reset and re-set BOTH λ=0.8 and μ=1; predict q(2)=3.6 and drain at t=20. Edits retain t and recompute the original fluid trajectory. Its initial four arbitrary audio units are not a recommended conversation buffer. The model has no text, prosody, vocoder, bursts, or listening score; it only illustrates why sustained production and consumption rates matter.