Customizing speech synthesis

Non-pair tags, also known as empty tags or self-closing tags, consist of a single tag that does not surround any content.

Non-pair tags are used to indicate that an element does not have any content but may have attributes. Non-pair tags end with a forward slash (/) before the closing angle bracket (>).

  ... <break time="2s"/> ...

This is part before the break <break time="50s"/> and this is part after the break

<emphasis> allows to emphasize a particular word or phrase in the speech. The l. .evel attribute specifies the level of emphasis, which can be either "strong" or "moderate".

<prosody> allows to adjust the pronunciation, volume, and speaking rate of the speech. The pitch, range, rate, and volume attributes can be used to adjust these aspects of the speech.

<say-as> allows to specify how a particular string of text should be pronounced. The interpret-as attribute specifies the type of text to be interpreted, and the format attribute specifies the format of the text. The detail attribute can be used to provide additional information about how the text should be pronounced.

<phoneme> allows to specify the pronunciation of a particular phoneme. The alphabet attribute specifies the phonetic alphabet being used, and the ph attribute specifies the phoneme to be pronounced.

<sub> allows you to substitute one word or phrase for another. The alias attribute specifies the replacement text, and the content between the start and end sub-tag represents the text to be replaced.

• In the dialogue window, you can see the width of each word segment(x-axis) on the intonation curve.

• Keep in mind that longer words contain more vowels and therefore more opportunities for the intonation curve to be modified.

• By default, the intonation curve is represented by a straight line at 0%. However, this does not mean that the basic intonation is completely flat! The tool allows you to adjust the pitch relative to the automatic basic synthesis.

• A maximum of five intonation points can be plotted on the curve for a single section.

A falling intonation pattern involves placing anchor points at the beginning of a phrase or sentence and gradually lowering the pitch as you move forward in time. This pattern often signals a statement or completion.

You can create a fluctuating intonation pattern by adding anchor points at various positions along the Intonation curve. This pattern involves both rising and falling pitch movements, conveying emphasis or highlighting important information.

A plateau intonation pattern maintains a relatively steady pitch without significant changes. You would position anchor points at similar heights along the curve, resulting in a flat or level pitch. This pattern is often used for conveying neutral or declarative statements.

A normal punctuation dot is an interruption of approximately 200-300 ms. If you find these pauses too long, we recommend replacing the punctuation with a shorter break. Ideally 100 ms, but never less than 50 ms.

By setting different pause lengths, we can influence the rhythm of speech. For example, set a shorter pause between sentences that form a single thought than at the front of different thoughts. Millisecond differences in pause length are almost imperceptible to the conscious human. However the different impression of synthesis is mainly created by the change of rhythm between sections. We can take advantage of this and better separate the individual information.

In the example shown in the figure above, the texting can be divided into the sections acknowledgement → argumentation → call to action → confirmation question. The argumentation consists of several sentences that form a unified train of thought, so it contains shorter pauses. On the other hand, a 100ms gap between sections is more distinctive.

The break also affects the intonation of the sentence and acts as an intonation dot, i.e. the intonation curve drops at the end.

Sometimes, however, we need to edit a pause between words or parts of sentences without intonating a full stop after the sentence. A softer drop in voice and at the same time a pause are provided by other signs, the most prominent of which is the punctuation mark, and slightly more subtle are the colon, semicolon or dash.

Prominent pauses: Něco končí <break strengh = "weak"/> Něco začíná. Něco končí <break time = "100ms" /> Něco začíná. Něco končí. Něco začíná.

Softer pauses: Něco končí; něco začíná. Něco končí, něco začíná Něco končí: něco začíná. Něco končí- něco začíná.

It is recommended that these marks be added at the point where the voice drop and short pause are to occur, or that they replace the original punctuation. If the pause is too short, write several of these signs in a row, adding their lengths together.

Něco končí,,, něco začíná. Něco končí;; něco začíná.

Example

Note that we have applied multiple strategies to improve the quality of the synthesis in the example.

1. We have rewritten the graphical distinction of the options with words (A. ---> option A);

2. Replacing punctuation marks with a break tag of a chosen length;

3. Combining additional punctuation marks between words for finer and shorter pauses with less effect on intonation.

The question mark serves as a punctuation mark that not only indicates a question but also functions as a pause. When a question is posed, there is typically a brief interruption of around 300 milliseconds.

However, if the question is followed by additional text, this pause can disrupt the flow of speech and give the impression that the voicebot has finished speaking. To prevent the voicebot and the user from overlapping in their speech, it is important to consider certain strategies for phrasing questions and placing them appropriately within the speech. The chapter on copywriting delves into these strategies in more detail.

Let's review some fundamental principles of good practice:

• Each message should contain only one question.

• Each question should focus on a single topic. For example, instead of asking, "Would you like to create a new account and join our Premium Club?" it is better to ask separate questions for each option.

• It is advisable to include the question at the end of the message so that the user has already received all the necessary information.

Let's address one specific case of choice questions (e.g., "Do you want A or do you want B?").

In this scenario, it is crucial to ensure good synthesis and emphasize distinctive intonation to convey the question's intent clearly and elicit the expected response. One recommendation is to structure the sentence in a way that presents each option as a separate sentence.

To assess the quality of synthesis, intonation, and the initial impression the question makes, experiment with different writing styles and punctuation. It is possible that breaks or punctuation marks may not provide a sufficiently distinct separation between the two options. In such cases, trying out a question mark after each option could be worth considering.

Please note that this is not a violation of the rule stating that only one question should be included in each text. It's important to remember that speech synthesis input notation serves a different purpose than regular text and does not necessarily adhere to grammar or spelling conventions. In this instance, it is considered a single question, with each option represented by a distinct intonation pattern that visually resembles a question mark.

In the case of "Do you want coffee? Or do you want tea?" a pause between the two options might be excessively long. Unfortunately, this gap cannot be shortened using the break tag. One possible approach is to remove the space after the first question mark (e.g., "Do you want coffee?Or do you want tea?"). In certain cases, combining the options into one continuous string can be helpful, with the question mark indicating the intonation but without the unwanted extended pause.

However, note that this is just one of the potential solutions and not the only correct way to adjust the intonation in such cases.

When incorporating English words into the 🇨🇿 Czech language, the stress patterns of these borrowed words tend to follow the stress patterns of Czech words. However, there may be some adjustments to fit the Czech stress rules. Here are a few guidelines for handling English words in Czech:

• First-syllable stress: As mentioned earlier, Czech generally has stress on the first syllable of a word. When adopting English words, the primary stress is often placed on the first syllable to align with Czech stress patterns.

Example: "computer" in English has stress on the second syllable (/kəmˈpjuːtər/), but in Czech, it would typically be pronounced with stress on the first syllable: /ˈkɔmpjutr/.

• Adaptation of vowel sounds: English vowels can have different qualities compared to Czech vowels. When adapting English words into Czech, the vowel sounds may be modified to match the Czech vowel inventory.

Example: "restaurant" in English (/ˈrɛstərɒnt/) could be adapted in Czech as /ˈrɛstaurant/.

• Retaining original stress: In some cases, English words may retain their original stress patterns, particularly when they are relatively recent borrowings or specialized terms that have become familiar to Czech speakers.

Example: "hotel" in English has stress on the first syllable (/hoʊˈtɛl/), and this stress pattern is often preserved when using the word in Czech: /ˈhotɛl/.

👩‍🏫 Adaptation of English words in Czech can vary depending on individual preferences, language register, and the familiarity of the borrowed word to Czech speakers. Therefore, there can be some variability in how English words are pronounced within the Czech language context.

19,99 - [devatenáct celých, devadesát devět setin] 19,99 Kč - [devatenáct korun devadesát devět haléřů] v akci jen za 19,99 Kč - [v akci jen za devatenáct devadesát devět] rohlík stojí 1,50,- - [rohlík stojí korunu padesát]

Azure's speech synthesis recognizes some words or signs for currency

A. Czech currency 100 Kč [sto českých korun] 1000 CZK [tisíc českých korun]

B. Dollars 10 $ [10 dolarů] 3 $ [3 dolary]

C. Euros 3 € [3 eura] 1000 € [tisíc eur] 1 € [jedno euro]

D. British pounds. 1 £ [jedna libra] 3 £ [tři libry] 10 £ [deset liber]

E. Not all international currency symbol are supported (besides few most common they're usually not)

Pronunciation of digits depends on:

• length of a digit string

• input format

• language and neural persona

• SSML rules

• reading rules and their localization

• numeric type (integer, float, ordinal)

🇸🇰 Slovak language

• If digits have fewer than or exactly 6 digits, they are always read decadically as default, regardless of whether a space separates orders of thousands.

123456 123 456 Both numbers are pronounced as: stodvadsaťtri tisíc štyristo päťdesiatšesť

• If digits have 7 or more characters, they are read decadically only if the order of thousands is separated by a space. Numeric string notation without spaces defaults to reading each digit in turn.

1234567 is pronounced [jedna dva tri štyri päť šesť sedem] 1 234 567 is pronounced [jeden milión dvestotridsaťštyri tisíc päťsto šesťdesiatsedem]

• The maximal length of spaced numerical string which is pronounced decadically is 15 digits (10^14), in EN system for hundreds of trillions, in SK stovky bilionov.

111 222 333 444 555 is pronounced [sto jedenásť biliónov ...]

• Spaced numerical strings longer than 15 digits are pronounced as multiple numbers, each containing 15 digits of less

111 222 333 444 555 666 will be divided as 111 222 333 444 555 | 666 [sto jedenásť biliónov ... päťstopäťdesiatpäť | šesťstopäťdesiatšesť

• If we need shorter numbers to be read each digit in turn, there are several ways to do it

Add spaces in between 1 2 3 4 5 6 is pronounced jedna dva tri štyri päť šesť

Add commas in between 1, 2, 3, 4, 5, 6 is pronounced jedna dva tri štyri päť šesť with more distinct pauses between each of them

Use SSML says-as for spelling (hláskování) <say-as interpret-as="spell" format="undefined">123456</say-as> is pronounced jedna dva tri štyri päť šesť

• If we need longer number to be read each digit in turn, copywriting input needs to be adapted accordingly

Use numeric string without spaces 1234567890

Separate each digit with spaces 1 2 3 4 5 6 7 8 9

Separated each digit with interpunction 1,2,3,4,5,6,7,8,9

These number will be pronounced as jedna dva tri štyri päť šesť sedem osem deväť

🇬🇧 English language

• If digits have fewer than or exactly 6 digits, they are always read decadically as default, regardless of whether orders of thousands are separated by a space.

1234 1 234 Both numbers are pronounced as: [one thousand two hundred thirty-four] 12345 12 345 Both numbers are pronounced as: [twelve thousands three hundred and forty-five] 123456 123 456 Both numbers are pronounced as:[one hundred twenty-three thousands four hundred fifty-six]

• If digits have 7 or more characters, they are read decadically only if the order of thousands is separated by a space. Numeric string notation without spaces defaults to reading each digit in turn.

1234567 is pronounced [one two three four five six seven] 1 234 567 is pronounced [one million two hundred twenty-three thousands five hundred sixty-seven]

• The maximum length of spaced numerical string which is pronounced decadically is 15 digits (10^14), in EN system for hundreds of trillions.

111 222 333 444 555 is pronounced [ one hundred and eleven trillion...]

• Spaced numerical strings longer than 15 digits are pronounced as multiple numbers, each containing 15 digits of less

111 222 333 444 555 666 will be divided as 111 222 333 444 555 | 666 [one hundred and eleven trillion ... five hundred fifty five | six hundred sixty six]

• If we need shorter numbers to be read each digit in turn, there are several ways to do it

Add spaces in between 1 2 3 4 5 6 is pronounced [one two three four five six]

Add commas in between 1, 2, 3, 4, 5, 6 is pronounced one two three four five six with more distinct pauses between each of them

Use SSML alias for spelling (hláskování) <say-as interpret-as="spell" format="undefined">123456</say-as> is pronounced one two three four five six

• If we need longer number to be read each digit in turn, copywriting input needs to be adapted accordingly

Use numeric string without spaces 1234567890 [one two three four five six seven eight nine zero]

If all subsequent triplets are made with 000, the string is read decadically even without spaces

• 100000 - [one million]

• 1100000 - [eleven million]

• 111000000 - [one hundred eleven million]

• 2000000000 - [two billion]

• 123000000000000 - [one hundred twenty-three trillion]

But when zeros cannot be divided into triplets, the numerical string is pronounced one digit at a time

• 123100000000000 - [one two three one zero zero zero ... ]

• If we need longer numbers to be read decadically, divide orders of thousands, millions, billions, trillions by space

• 1 234 567 - [one million to hundred thirty-four thousands five hundred sixty-seven]

• 123 000 000 000 001 - [one hundred twenty-three trillion and one]

🇵🇱 Polish language

• If digits have exactly 4 digits.

A. If we have numbers 1000, 2000, 3000, etc. The numbers are read by Azure neural voices as: 1000 - tysięczny rok – year one thousand 2000 - dwutysięczny rok – year two thousand 3000 - trzytysięczny rok - year three thousand etc. - This is WRONG. It should be pronounced as: 1000 - tysiąc 2000 - dwa tysiące 3000 - trzy tysiące

This situation can also occur with other numbers, for example 4999 is pronounce as cztery tysiące dziewięćset dziewięćdziesiąty dziewiąty rok - this is WRONG. CORRECT is cztery tysiące dziewięćset dziewięćdziesiąt dziewięć. Every number needs to be checked!

B.

1..., 2..., 3..., 4... (thousand) are pronounced CORRECT.

5..., 6..., 7..., 8..., 9... (thousand) are pronounced WRONG.

5324 – pięć TYSIĄCE - correct is pięć TYSIĘCY

6234 - sześć TYSIĄCE - correct is sześć TYSIĘCY

7324 – siedem TYSIĄCY - correct is siedem TYSIĘCY

8324 - osiem TYSIĄCE - correct is osiem TYSIĘCY

9324 - dziewięć TYSIĄCE - correct is dziewięć TYSIĘCY

🇸🇰 Slovak language

For the pronunciation of telephone numbers:

A. Write them down in iso format including prefix Volajte +421 800 148 148 is pronounced VOLAJTE PLUS ŠTYRI DVA JEDEN, OSEMSTO, STOŠTYRIDSAŤOSEM, STOŠTYRIDSAŤOSEM

B. Separate custom sections with interpunction Volejte 800, 148, 148 is pronounced OSEMSTO, STOŠTYRIDSAŤOSEM, STOŠTYRIDSAŤOSEM Volejte 800-148-148 is pronounced OSEMSTO, STOŠTYRIDSAŤOSEM, STOŠTYRIDSAŤOSEM Volejte 800, 54, 12, 12 is pronounced OSEMSTO, PÄŤDESIATŠTYRI, DVANÁSŤ, DVANÁSŤ Volejte 800, 12, 7, 7, 7. 7 is pronounced OSEMSTO, DVANÁSŤ, SEDEM, SEDEM, SEDEM, SEDEM

C. Write them down as alphabetical string Volajte na osemsto dvanásť štyri siedmičky Volajte na osemsto dvanásť sedem sedem sedem sedem

🇬🇧 English language

For the pronunciation of telephone numbers:

A. Write them down in iso format including prefix Call +1-212-456-7890 (USA) is pronounced [call plus one - two one two - four five six - seven eight nine zero] Call +44 7911 123456 (UK) is pronounced [call plus four four, seven nine one one, one two three four five six]

B. For USA, obey domestic 3-3-4 format Call 212-456-7890 is pronounced [call one - two one two - four five six - seven eight nine zero]

C. For the UK, obey area code formatting ( 3-4 digits + 8-7 digits) 020 (London) 1234 5676 is pronounced [zero two zero, one two three four, five six seven eight]

D. To customize pronunciation, separate digits with spaces or punctuation Call 4 4 4 4, 1 2 3, 1 2 3 is pronounced [call four four four four, one two three, one two three] with more distinctive pauses between comma-separated sections Call 800, 12, 34, 12, 34 is pronounced [call eight hundred, twelve, thirty- four, twelve, thirty-four] Call 800 12 34 12 34 is pronounced [call eight hundred one two three four one two thirty-four

E. For 10-digit phone number (prefix not included), you can use SSML alias with attribute <say-as interpret-as="telephone" format="undefined">1234567890</say-as> [one two three four five six seven eight nine o]

🇸🇰 Slovak language

For pronunciation of long numerical strings (order IDs etc.):

A. Write numerical strings without spaces to be read one digit in turn Objednávka 01304578931 Objednávka NULA JEDEN TRI NULA ŠTYRY PÄŤ SEDEM OSEM DEVÄŤ TRI JEDEN

B. Customize pronunciation by breaking it into smaller chunks with interpunction Objednávka 01, 30, 45, 78, 9-3-1 Objednávka NULA JEDEN, TRIDSAŤ, ŠTYRIDSAŤPÄŤ, SEDEMDESIAŤOSEM, DEVÄŤ TRI JEDEN

🇬🇧 English language

For pronunciation of long numerical strings (order IDs etc.):

A. Write numerical strings without spaces to be read one digit in turn Order 01304578931 [o one three o four five seven eight one three one]

B. Customize pronunciation by breaking it into smaller chunks with interpunction Order 01, 30, 45, 78, 9-3-1 [zero one, thirty, forty-five, seventy-eight, nine three one]

🇸🇰 Slovak language

Dates are pronounced correctly by default when put down in ISO format

A. ISO DD-MM-YYY 01-11-1995 1-11-1995 Pronounced as [1. novembra 1995]

B. ISO YYYY-MM-DD 1995-11-01 1995-11-1 Pronounced as [1. novembra 1995]

C. ISO DD.MM.YYY datum 01.11.1995 datum 01. 11. 1995 datum 1.11.1995 datum 1. 11. 1995 Pronounced [as 1. novembra 1995]

Safe and simple way is to transcribe date alphanumerically

dátum 1. novembra 1995 prvého novembra 1995 Pronounced as [1. novembra 1995]

prvého jedenástý 1995 Pronounced as [prvého jedenástý 1995]

These formats listed below WON'T WORK and will be pronounced incorrectly, even if tagged with SSML alias reading rules for dates

• DD.MM - 1.11. - [jedna hodina jedenásť minút]

• DD. MM - 1. 11. - [prvý jedenásť]

• DD/MM - 01/11 - [nula jeden lomka jedenásť]

• DD/MM/YYYY - 01/11/1995, 1/11/1995 - [jeden lomka jedenásť lomka 1995]

• SSML alias reading rules not supported

🇬🇧 English language

Dates are pronounced correctly by default when put down in ISO format

A. ISO DD-MM-YYY 01-11-1995 1-11-1995 Pronounced as [the first of November 1995]

B. ISO YYYY-MM-DD 1995-11-01 1995-11-1 UK and US: Pronounced as [the first of November 1995]

C. ISO DD.MM.YYY date 01.11.1995 datum 1.11.1995 UK: Pronounced as [the first of November 1995] US: Pronounced as [January eleventh 1995]

D. ISO DD/MM/YYYY and YYYY/MM/DD 01/11/1995 1995/11/01 UK: Pronounced as [the first of November 1995] UK: Pronounced as [Janualy eleventh 1995]

D. Also Nov,1st 1995 November 1st 1995 Pronounced as [November first, 1995] 1 November 1995 1 Nov 1995 1st November 1995 Pronounced as [the fist of November 1995]

These inputs WON'T BE pronounced correctly

• 1 st November 1995 - [one es tee November 1995]

• date 1. November 1995 - [one November 1995]

• November, 1. 1995 - Pronounced as [November one 1995]

• 01/11 (EN-UK voices) - [zero one eleventh]

• 1.11.1995 - [one. eleven. 1995]

• Be mindful of US date format MM/DD

• 01/11 (EN-US voices) - [January the eleventh]

• Various SSML tag says-as formats of attribute date are supported

• DMY - US:[November 1st 1995] , UK:[the first of November 1995]

• MDY: <say-as interpret-as="date" format="mdy">1/11/1995</say-as>- US:[January 11th 1995] , UK:[the 11th of January 1995]

• MD: <say-as interpret-as="date" format="md">1/11</say-as> - US:[January 11th] , UK:[the 11th of January]

• DM: <say-as interpret-as="date" format="dm">1/11</say-as>- US:[November 1st] , UK:[the 1st of November]

• MY: <say-as interpret-as="date" format="my">11/1995</say-as>- US, UK:[November 1995]¨

• YM: <say-as interpret-as="date" format="ym">1995/11</say-as>- US, UK:[November 1995]

🇸🇰 Slovak language

Standard iso format is supported and is pronounced as time correctly by default

• HH:mm - 16:10 - [šesťnásť hodín desať minút]

• HH:mm:ss - 16:10:10 - [šesťnásť hodín desať minút desať sekúnd]

• HH.mm - 16.10 - [šestnásť hodín desať minút]

Whole hours, even if written down digitally, are pronounced as analog time

• HH:00 - 16:00 - [šestnásť hodín]

• HH:00:00 - 16:00:00 - [šestnásť hodín]

Slovak SI unit system IS SUPPORTED only for hod and min

16 hod - [16 hodín] 15 min - [15 minút] 15 h 10 min - [15 hodín 10 minút]

Won't read properly other SI formats such as 16 h - [16 h] 10 s - [15 s] 10 sek - [10 sek]

If you need word seconds to be pronounce, write that down manually 16 hod 10 min 10 sekúnd - [16 hodín 10 minút 10 sekúnd]

In case you need time to be pronounced as analog or Slovak SI units to be pronounced correctly, it is necessary to transcribe your input

• 15:15 → štvrť na štyri

• 12:00 → poludnie

• 12:30 → pôl jednej

• 10 s → 10 sekúnd

• po 8:00 → po osmej hodine

🇬🇧 English language

Standard ISO format is supported and pronounced correctly by default:

• 15:15 → quarter past three

• 12:00 → at noon

• 12:30 → half past twelve

• 10 s → 10 seconds

• around 8 - around 8-ish

In case you need time to be pronounced as analog or time notation to be pronounced correctly, it is necessary to transcribe your input:

A. 10 o'clock [10 o'clock]

B. 10:00 AM 10 a.m. 10 AM [10 A. M.]

Also, other formats of time are supported:

A. HH:mm:ss 10:50:10 [10 hours 5 minutes and 20 seconds]

B. HH:mm 15:00 UK,US: [three P.M.] 10:00 UK, US: [ten o'clock] 10:45 UK, US: [ten forty five]

✅ Standard ISO format is supported and pronounced correctly by default.