Kojo’s speech exceeds the synthesis limit. His cousin needs the audio before morning.

Long narration is divided at sentence boundaries into requests that fit the configured synthesis limit, then the resulting WAV chunks are recombined in order. Each request must remain within that limit, and incompatible chunk audio causes an explicit failure instead of producing a damaged file.

Consider Kojo, an invented composite, standing in his kitchen in Kumasi at 9:40 p.m. He is holding a marked-up speech for his aunt’s retirement celebration, with Twi greetings, Ghanaian English, and several paragraphs of family memories. The recording needs to reach his cousin before morning so she can rehearse it on her drive.

Kojo pastes the whole narration into one synthesis request. It is too long. Cutting the text carelessly could split a sentence halfway through, leaving one audio file ending on “because” and the next beginning after an unnatural pause.

If he cannot prepare a usable recording tonight, his cousin will rehearse from the page alone. The Twi names and shifts between languages are the parts she asked to hear.

Why sentence boundaries matter

A long passage has to become smaller requests somewhere. Sentence-aware chunking chooses boundaries that already belong in spoken language, such as the end of a statement or question.

That choice preserves meaning and rhythm better than chopping at a fixed character position. Compare a split after “We thought the rain had ended” with a split after “We thought the rain had.” Both may satisfy a technical size limit, but only one gives the voice a sensible place to stop.

The chunker works within a hard constraint: every resulting request must fit the configured limit. A single sentence that exceeds that limit still needs another valid split. Sentence awareness helps choose cleaner boundaries; it does not remove the limit.

This matters even more when the narration moves between Twi and Ghanaian English. The text should remain natural before synthesis begins. If you are preparing a mixed-language message, this rehearsal example shows why hearing the full phrasing can expose trouble that silent reading misses.

From one script to several WAV chunks

Kojo makes a working copy of his speech and checks it paragraph by paragraph. He removes a repeated introduction, fixes an unfinished sentence, and keeps each language change where it belongs in the thought.

The narration can then follow a practical sequence:

1. Start with clean text whose sentences end clearly.
2. Divide it at sentence boundaries while keeping every request under the configured limit.
3. Synthesize each chunk in its original order.
4. Retry a chunk when a temporary failure prevents a valid result.
5. Verify that the completed audio chunks are compatible.
6. Recombine the WAV audio in sequence.

Retries apply to the failed chunk, so a temporary error does not require treating the entire narration as one indivisible job. A retry still has to return usable audio. Repeatedly sending a malformed or oversized request will not repair the underlying problem.

Order matters too. Labeling chunks by sequence, rather than by whatever filename happens to be generated, prevents paragraph four from appearing before paragraph three. For a speech, lesson, or family message, that simple discipline protects the story.

Compatibility protects the final recording

WAV files can share the same extension while differing in properties required for valid recombination. Joining incompatible chunks as raw bytes can create a file that plays incorrectly, stops early, or contains misleading header information.

The safe behavior is explicit: incompatible chunk audio fails rather than being silently joined. That visible failure tells you where to investigate. Check whether the chunks were produced with matching audio settings, replace the incompatible result, then attempt recombination again.

This follows the same principle as showing a synthesis error instead of swallowing it. A clear failure may interrupt the task, but it protects the listener from receiving an audio file that only appears complete.

Text preparation also deserves attention. If the intended language and the supplied speech text conflict, splitting the passage perfectly will preserve the conflict across several chunks. Locale and speech text must agree before synthesis starts.

Prepare the script before pressing record

With the deadline close, Kojo stops trying to force the whole speech through at once. He checks the longest sentences, gives each paragraph a clear ending, and processes the narration as ordered chunks. When one chunk fails, he corrects that part instead of guessing whether the full recording worked.

The compatible WAV chunks are then recombined in sequence. His cousin receives one continuous narration, with the retirement greeting first, the family memory in the middle, and the closing message where it belongs.

Before preparing your own long narration, read it aloud once. Shorten any sentence that is difficult to finish in one breath, confirm every chunk can fit the configured limit, and keep the sequence visible from the first request to the final WAV.