AI Audiobook Narration Reads “银行” as “行走”: How Much Can a Pronunciation Lexicon Actually Fix in 2026?

AI audiobook narration reading “银行” (bank) as “行走” (walking), or reading “2026年” as “二零二六年,” is in most 2026 projects a text-layer problem, not a sign that the speech model is not strong enough. In practice, 60–80% of deterministic pronunciation errors can be reduced across the three layers of text normalization, polyphone lexicons, and pronunciation markup; switching models mainly improves timbre and prosody, with limited help for pronunciation disambiguation. Which layer to fix first depends on your text type and acceptance criteria.

Why a stronger speech model often does not fix mispronunciations

A common speech synthesis pipeline is: text normalization (converting numbers, units, symbols, and English abbreviations into readable forms) → word segmentation and polyphone disambiguation → prosody and phrasing → acoustic model generation → vocoder waveform reconstruction. Mispronunciations mostly occur in the first two steps, while model upgrades usually change the last two, which is why the same mispronounced positions remain after switching models.

The test is not complicated: align two audio versions of the same text. If the mispronounced positions match, and they still match after switching to another model provider, the problem is basically upstream in text processing. At that point, do not rush to spend budget on a model change.

• Polyphonic characters: characters such as 行, 重, 长, 了, 得, 都, 还, 传, 差, 发 need context to determine their pronunciation.
• Numbers and units: years, decimals, percentages, amounts, phone numbers, and serial numbers are prone to problems when reading rules are inconsistent.
• English and abbreviations: whether to read by letters or as words often depends on industry convention; do not leave it to the model’s default handling.
• Proper nouns: personal names, place names, book titles, and brand names are the part where lexicon coverage gives higher returns.
• Tone and neutral tone: erhua, neutral tone, and function words affect whether it sounds human, not whether it is correct.

A four-layer funnel for fixing mispronunciations

It is called a “four-layer funnel” because misreadings leak out one layer at a time along the same pipeline. The earlier you fix, the lower the cost and the more reusable the result; the later you fix, the higher the cost, because reworking generated audio often means rerunning whole segments. So the order is not arbitrary: do the intake first, then markup, and only then talk about sampling and feedback.

1. Intake normalization: Convert numbers, units, symbols, and English according to rules. The rule table must be configurable, and keep both the source text and the narration text so you can roll back and reconcile. Do not overwrite the source text directly; otherwise, when the client changes it back, there is no way to check.
2. Polyphone disambiguation: Use a lexicon plus context rules, prioritizing high-frequency proper nouns and industry terms. One manual confirmation can be reused long term. The lexicon needs version numbers; with multiple collaborators, no versioning easily leads to overwrites.
3. Pronunciation and prosody markup: Use SSML or platform-specific tags to control pauses, stress, and character-by-character pronunciation, marking only error-prone segments. Over-marking makes the tone stiff; it is better to mark a few fewer spots.
4. Sampling and feedback: After launch, spot-listen by percentage and feed mispronounced segments back into the lexicon to form a versioned word bank. The key is whether you can locate and rerun by sentence, not rerecord the whole book.

Expand the lexicon or switch models: an experience-range comparison

These two actions solve problems at different levels; comparing them side by side makes the account clearer. Below is a comparison based on experience ranges from typical 2026 projects. Specific numbers vary with text volume, voice, and platform billing.

• Time to see effect: Lexicon plus normalization usually covers high-frequency errors in a few days to two or three weeks; switching models involves integration, listening tests, and acceptance, typically one to four weeks, and it does not guarantee better pronunciation.
• Cost structure: Lexicons and normalization are mainly labor; the experience range is from a few thousand to tens of thousands of RMB, depending on word count and proper-noun density. If switching models changes both voice and billing, the whole audio batch may need regeneration; in practice, incremental cost may be 20–50% of the original budget.
• What it can fix: A lexicon can fix deterministic, enumerable misreadings; switching models can improve timbre naturalness, breathiness, and emotional expression.
• What it cannot fix: A lexicon cannot fix stiff prosody and mechanical feel; switching models cannot fix proper nouns and industry terms not in the lexicon.
• Best fit: For batch long-form text with many proper nouns, do the lexicon first; for short spoken lines where you are unhappy with timbre and emotion, then consider switching models.

One easily overlooked boundary: if all misreadings are concentrated in one uncommon industry term, a proper-noun lexicon of a few dozen lines is enough. At that point, moving to a larger model instead brings new billing and review headaches.

On delivery: store source text and narration text separately first

In audiobook or course voiceover delivered according to enterprise project habits, the client often gets stuck at “give us three chapters for a listening test first”: budget and time only allow one pass, while source materials are often scanned PDFs or Word files with comments, and number and symbol formats are inconsistent. Our approach is to first store the source text and narration text of the sample chapters separately, split by paragraph, mark error-prone segments individually, and do a manual listen before finalizing. The trade-off is one to two extra days for splitting and markup on the three sample chapters; the benefit is that the formal batch does not need a full-book rerun, typically saving one full-volume rework, and the lexicon can carry over directly to later chapters.

Acceptance criteria and common pitfalls

Do not use “it sounds okay” as acceptance criteria. Recommended: spot-listen by paragraph, starting with 5% to 10%, focusing on proper nouns, numbers, and dialogue paragraphs; record mispronunciations per 10,000 words. A typical acceptable range is single digits to a few dozen per delivery. List the remaining misreadings for the content owner to confirm, rather than saying verbally that it is fine. This criteria applies to both internal acceptance and client acceptance.

• Testing only with short sentences; long-sentence phrasing and pauses are not validated and only show up at scale.
• Ignoring number and unit rules; amount readings are found wrong only during listening.
• No version control for the lexicon; several people edit it and overwrite each other the next day.
• Generated audio not stored by sentence; changing one sentence requires rerunning a whole chapter, multiplying rework cost.
• Ignoring platform content review requirements and voice authorization scope, leading to takedowns after launch.

Against these pitfalls, what counts as qualified? At least three things: source text and narration text can be compared and rolled back; audio can be located and rerun by paragraph or sentence; mispronunciation lists are recorded and can be fed back into the lexicon.

Where it fits and where it does not

Good fit: audiobooks, course and knowledge-payment voiceover, batch short-video narration, internal corporate training audio, and base-track voiceover for multilingual versions. This content has large text volume, many proper nouns, and relatively moderate emotional performance demands, so the four-layer funnel gives clear returns.

There are also cases where this approach is not suitable or unnecessary: radio drama character voiceover requiring nuanced emotional performance is still better with humans primarily and AI assisting; real-time dialogue scenes under a few dozen characters are easier with human fallback than building a lexicon; publication-level zero-misreading requirements can only be made less likely by lexicons and sampling, and still need human proof-listening. Distinguishing these cases saves more budget than adding another model layer.

FAQ

In AI voiceover, how many polyphonic-character misreadings can a pronunciation dictionary fix?

In practice, 60–80% of high-frequency proper nouns, industry terms, and common polyphonic characters can be fixed; the rest relies on context rules and human proof-listening.

Will a newer speech model make polyphonic-character problems disappear automatically?

Generally no. Pronunciation disambiguation mostly happens in the upstream text layer. Switching models mainly changes timbre and prosody, and the mispronounced positions often remain.

For an audiobook of hundreds of thousands of words, what sampling ratio is appropriate?

Typically start from 5% to 10%, focusing on proper nouns, numbers, and dialogue paragraphs; record mispronunciations per 10,000 words, then define the rework scope.

What is the cost difference between cloud synthesis and maintaining your own pronunciation lexicon?

Cloud synthesis is usage-billed; for hundreds of thousands of words, the experience range is often a few hundred to a few thousand RMB. Self-hosting requires server and labor costs, usually evaluated after call volume stabilizes.

What should be noted about voice cloning and voiceover authorization?

Use voice sources with clear authorization, keep authorization evidence and scope of use, and check commercial scenarios against platform rules and contract terms. Do not clone online material directly.

If you are stuck on acceptance for polyphonic characters and number misreadings, first fill in text normalization and the lexicon using the four-layer funnel, run three to five sample chapters before finalizing, and save budget for what the model layer can actually improve, such as timbre and emotion. Conversely, for radio dramas requiring nuanced performance or publication-level zero-misreading content, do not expect the lexicon to cover everything; still arrange human proof-listening or even human recording.