A museum voiceover script includes an ambiguous date, a catalog accession number, and a measurement on a historic display. Text-to-speech can say all three fluently while still confusing a listener. Amazon Polly supports the SSML say-as element to indicate how selected text should be interpreted, including date, digits, cardinal numbers, and units. The source text remains the archival authority; rendered speech is an accessible derivative. This guide uses invented collection details, not a recording from a real museum or a claim that we synthesized audio.
Treat the archival record as a fixed source
Start with a curator-approved display label and record its revision before writing SSML. A painted date such as 04-05-1910 could represent April 5 or May 4. The correct interpretation must come from the institution’s catalog, not an AI guess about what date feels plausible. Put the collection identifier, locale, original transcription and approved spoken expansion in separate fields. That allows later corrections to be traced rather than silently editing the archival fact.
Ask who owns the pronunciation decisions. For a public exhibition, a curator can confirm dates, a language specialist can check spoken place names, and an accessibility reviewer can judge intelligibility. A text-to-speech request is not an approval mechanism. Keep a human-readable pronunciation note that survives a provider change, plus a neutral version of the copy that a screen reader can read without interpreting raw SSML markup.
Use say-as for interpretation, not source correction
AWS documents the SSML say-as element with an interpret-as attribute and, for dates, an optional format rule such as mdy, dmy, or ymd. Digits can be pronounced one by one when a code should not become a large cardinal number. For a sample accession ID such as 0427, digits make more sense than four hundred twenty-seven. If the historical text contains a wrong number, SSML must not quietly repair it; the institution must publish a separate correction.
In an implementation, escape XML-sensitive characters from approved prose and construct the SSML with a validated template. Surround the content with speak tags, and test the precise engine and voice used for delivery. Do not rely on a prompt or language model to generate unrestricted SSML for a production feed. Store the SSML alongside the plain-text script with revision history so a future editor can see which spans were deliberately expanded.
Check engine support before choosing characters
The say-as modes are not equally supported across Polly engines. AWS notes that when a neural voice encounters the characters or spell-out option, the affected sentence is synthesized with the related standard voice, even though neural-voice billing applies. That can create a noticeable timbre change inside an otherwise consistent museum narration. For identifiers, test digits instead of characters when the content is numeric, and confirm the result by listening to that exact voice.
The supported-tags table also distinguishes standard, neural, long-form and generative availability. Avoid assuming that a sample copied from a different engine or language works unchanged. The telephone option, for example, has documented language restrictions and is not a generic identifier formatter. Place a short synthetic example with an accession number, date and unit in a test script; never submit private donor notes merely to confirm basic SSML behavior.
Design the four-stage pronunciation handoff
The process diagram moves from an approved written date, to a narrowly tagged spoken form, to a voice-engine compatibility check, and finally to a listening review. The example format mdy is intentionally shown as a choice that requires evidence from the source. A markup parser can accept a syntactically valid date while still conveying the wrong day and month. The reviewer should compare the read-aloud version with the correct catalog date, not just with the literal digits.
Choose one primary engine and voice for each language edition. Keep an explicit fallback if a tag is unsupported or a standard-voice substitution sounds disruptive. A fallback may be a differently worded approved plain-text sentence rather than a tag. Document that choice as an editorial change; do not disguise it as an automatic improvement in vocal accuracy. Verify the surrounding safety message remains in the same order.
Compare three genuinely different narration states
Imagine an exhibit date written as 05-04-1910. The first state is an unmarked ambiguous string. The second is tagged for an approved date order and illustrates the intended month and day. The third is a reviewed public audio line paired with a matching on-screen date. These are proposed editorial states, not measured audio outputs or screenshots of Amazon Polly. The graphic should show the changed interpretation, not pretend a green check proves historical truth.
A second distinction is an object code such as 0042. If the code is spoken as forty-two, the leading zeros disappear, which may confuse people finding the object later. If read as zero zero four two, it preserves its lookup structure. Whether the museum wants letters spelled or a nickname pronounced depends on its catalog policy. Capture a pronunciation checklist for dates, leading zeros, foreign names, and units rather than making all numerals follow one rule.
Keep alternatives and accessibility available
An audio guide should be usable without access to a premium synthesis voice. Supply a faithful readable transcript, meaningful captions where relevant, appropriate volume, and a way to repeat an important date. For bilingual exhibits, do not copy English-specific date order into another locale. A language expert should decide whether units, calendars and digits are naturally spoken in that language while maintaining the content of the approved label.
Avoid using expletive-style beeps as a general redaction or accessibility tool; silence or beeps can remove meaning in a historical quotation. Where a curator has authorized an omission, label it transparently in both text and speech. An automatic audio export can also be cached by an app; version the media file and its text together so the kiosk does not play an old correction after the label has been revised.
Audit the places where pronunciation can fail
Build a small test set of invented date formats, leading-zero codes, fractions, measurements, and mixed alphanumeric identifiers. Check the engine response for invalid SSML and the delivered audio for unintuitive stress, pauses or voice switching. Include a test where date order is deliberately reversed. Successful synthesis does not mean the pronunciation conveys the correct archival fact. A listener unfamiliar with the exhibit should be able to repeat the intended date and code.
Record the selected engine, voice, SSML revision, display-text revision, reviewer and approval date. If a voice or engine changes, rerun the listening checks. Costs and available voices can change, so treat the service’s current documentation as authoritative at deployment. A caption comparison, independent of audio, should confirm that the visual text and spoken meaning match, especially for emergency directions or collection loan restrictions.
Release a documented speech edition
Approve the audio only after the exhibit owner confirms each source fact and a reviewer listens for unwanted substitution or changed date order. Retain the approved plain-text text, SSML, export hash and rollback instructions. Offer staff a direct correction route when a visitor notices that a spoken code and the visible label differ. The museum can update an audio edition without changing its historic catalog record, but it must explain that derivative edition to maintain provenance.
The useful outcome of say-as is a constrained interpretation rule, not a guarantee of natural performance. AWS supplies syntax and engine rules; institutional staff supply evidence and sign-off. For a small exhibition with just two ambiguous dates, writing an explicit approved spoken sentence may be clearer and cheaper than maintaining elaborate SSML. Choose whichever approach remains understandable, reviewable and compatible with the published text.
How We Research
For our November 2026 editorial review, we consulted the named official developer sources: AWS say-as tag documentation; Amazon Polly supported SSML tags; Amazon Polly SSML limits. We compared documented feature scope, request shapes, output limitations, and failure cases with a clearly hypothetical nonproduction museum or community-history scenario. We did not run paid API calls, process private recordings, access donor or visitor records, benchmark vendors, measure recognition accuracy, or capture actual product screenshots. All illustrations are original explanatory drawings rather than real output or proof of performance. Prices, availability, privacy terms, model support, and preview features can change; check the current official documentation and an authorized local test before deploying.
Frequently Asked Questions
No. It changes how text is spoken; source corrections require a separate curator-approved editorial action.
No. AWS documents a standard-voice substitution for the characters option with a neural voice.
Consider digits so each digit can be heard, then verify the sound and the museum’s approved code format.
No. Date order and speech conventions depend on locale, voice and institutional source evidence.
No. The delivered recording still needs human listening and comparison with approved visible text.