Шаблоны запросов TTS для естественной речи

Чем отличается составление запросов для TTS

Длина предложений: чем короче, тем лучше

# Prompt a model to generate TTS-optimized text
TTS_SYSTEM_PROMPT = (
    'You are writing text that will be read aloud by a text-to-speech system.\n\n'
    'Rules:\n'
    '- Use short, clear sentences (10-20 words each).\n'
    '- Avoid complex nested clauses.\n'
    '- End each sentence with a period for clear pause signals.\n'
    '- Avoid parenthetical asides in the middle of sentences.\n'
    '- Use conversational vocabulary — write as you would speak.\n'
    '- Never use bullet points, headers, or markdown formatting.'
)

# Bad (long, nested):
BAD = (
    'The transformer architecture, which was introduced in the landmark 2017 paper '
    '"Attention is All You Need" by Vaswani et al. at Google, fundamentally changed '
    'natural language processing by replacing recurrence with self-attention.'
)

# Good (TTS-friendly):
GOOD = (
    'The transformer architecture changed natural language processing. '
    'It was introduced in 2017 by researchers at Google. '
    'Their key innovation was replacing recurrence with self-attention.'
)

Пунктуация как звуковые подсказки

TTS_PUNCTUATION_PROMPT = (
    'Write the following content for text-to-speech narration. '
    'Use only periods, commas, and question marks for punctuation. '
    'Avoid em dashes, semicolons, colons, and parentheses. '
    'Break complex thoughts into multiple short sentences.\n\n'
    'Content to rewrite: {content}'
)

# Example transformation
original = (
    'There are three main factors to consider: cost, quality, and speed — '
    'and often, you can only optimize for two of them (this is sometimes '
    'called the project management triangle).'
)

for_tts = (
    'There are three main factors to consider. The first is cost. '
    'The second is quality. The third is speed. '
    'Usually, you can only optimize for two of these at once. '
    'This is sometimes called the project management triangle.'
)
print(for_tts)

Как избегать неоднозначных сокращений

TTS_ABBREVIATION_RULES = (
    'When writing for text-to-speech:\n'
    '- Write "Doctor" not "Dr."\n'
    '- Write "Street" or "Saint" not "St."\n'
    '- Write "versus" not "vs."\n'
    '- Write "et cetera" not "etc."\n'
    '- Write "for example" not "e.g."\n'
    '- Write "that is" not "i.e."\n'
    '- Write out years: "twenty twenty-five" not "2025" when narrated in prose\n'
    '- Write out numbers under 10: "three" not "3" in conversational text\n'
    '- Spell out acronyms on first use: '
    '"Application Programming Interface, or API"\n'
)
print(TTS_ABBREVIATION_RULES)

Слова, вызывающие неправильное произношение

# Prompting LLM to detect and fix TTS pronunciation issues
TTS_REVIEW_PROMPT = (
    'Review the following text for text-to-speech pronunciation issues.\n'
    'Identify words that might be mispronounced by a TTS engine because they:\n'
    '1. Are homographs with multiple pronunciations\n'
    '2. Are technical terms, brand names, or foreign words\n'
    '3. Are abbreviations or acronyms\n'
    '4. Are numbers in unusual formats\n\n'
    'For each issue, provide the original text and a TTS-safe replacement.\n\n'
    'Text to review: {text}'
)

# Example issues and fixes
ISSUES = {
    'She read the docs': 'She read (past tense - ambiguous) -> She finished reading the docs',
    'Lead developer': 'Lead (metal or guide?) -> Lead (rhymes with feed) developer',
    'API v2.3.1': 'v2.3.1 -> version 2 point 3 point 1',
    'The .env file': 'dot E N V file (may say .env) -> the environment config file',
    'Re-init': 'may say Ray-in-it -> reinitialize',
}
for problem, fix in ISSUES.items():
    print(f'Issue: {problem}\nFix: {fix}\n')

Числа и даты для TTS

NUMBER_TTS_RULES = (
    'When writing numbers for TTS narration:\n'
    '- Dates: write "January fifteenth, 2025" not "01/15/2025"\n'
    '- Time: write "3 in the afternoon" not "15:00" in casual speech\n'
    '- Percentages: write "forty-five percent" not "45%" in narration\n'
    '- Large numbers: write "one point two million" not "1.2M"\n'
    '- Phone numbers: spell with hyphens "555-123-4567" (TTS reads each digit)\n'
    '- Currency: write "twelve dollars and fifty cents" not "$12.50" in speech\n'
    '- Fractions: write "three quarters" not "3/4"\n'
    '- Ordinals: write "first" not "1st" (TTS may say "one S T")\n'
)

# Apply when prompting the LLM to generate TTS scripts
TTS_SCRIPT_PROMPT = (
    'Write a 30-second product announcement for text-to-speech.\n'
    'Product: TaskFlow Pro, $49/month, 10,000 users worldwide, '
    'launched January 2025.\n\n'
    + NUMBER_TTS_RULES
)

Удаление визуального форматирования

TTS_FORMAT_CONVERSION_PROMPT = (
    'Convert the following markdown text into plain prose suitable '
    'for text-to-speech narration.\n\n'
    'Rules:\n'
    '- Remove all markdown formatting (**, *, #, -, etc.)\n'
    '- Convert bullet lists into spoken sequences '
    '("First... Second... Third...")\n'
    '- Convert headers into topic introduction sentences\n'
    '- Remove any hyperlinks or URLs\n'
    '- Convert code snippets into descriptions '
    '("a Python function that...")\n\n'
    'Markdown text:\n'
    '{markdown_text}'
)

EXAMPLE_MARKDOWN = (
    '## Getting Started\n'
    '- Install the package with 'pip install mylib'\n'
    '- Set your **API key** in '.env'\n'
    '- Run 'python main.py' to start\n'
)

EXAMPLE_TTS = (
    'To get started, install the package using pip. '
    'Next, set your API key in your environment configuration file. '
    'Finally, run the main Python script to start.'
)
print(EXAMPLE_TTS)

Управление темпом с помощью переходных фраз

TRANSITION_PHRASES = {
    'starting_new_topic':    ['Let us now talk about', 'Moving on to', 'Next,'],
    'adding_information':    ['Additionally,', 'Also worth noting,', 'Furthermore,'],
    'contrasting':           ['However,', 'On the other hand,', 'That said,'],
    'concluding':            ['To summarize,', 'In short,', 'To wrap up,'],
    'sequencing':            ['First,', 'Second,', 'Then,', 'Finally,'],
    'emphasizing':           ['Importantly,', 'Keep in mind that', 'Note that'],
}

TTS_STRUCTURE_PROMPT = (
    'Write a two-minute explainer on {topic} for text-to-speech narration.\n'
    'Use verbal transition phrases to guide listeners through each point. '
    'Avoid headers or bullet points. '
    'Structure the content with clear spoken signposts '
    '("First...", "Moving on...", "To summarize...").'
)

print('Available transitions:')
for category, phrases in TRANSITION_PHRASES.items():
    print(f'  {category}: {phrases[0]}')

Проверка текста для TTS

import openai
import pathlib

client = openai.OpenAI(api_key='sk-...')

def generate_and_test_tts(text, output_file='test_audio.mp3'):
    """Generate TTS audio from text for auditory review."""
    response = client.audio.speech.create(
        model='tts-1-hd',
        voice='alloy',   # alloy, echo, fable, onyx, nova, shimmer
        input=text,
        speed=1.0        # 0.25 to 4.0
    )

    audio_path = pathlib.Path(output_file)
    response.stream_to_file(audio_path)
    print(f'Audio saved to {audio_path}')
    print(f'Text length: {len(text)} chars, {len(text.split())} words')
    return audio_path

# Generate and listen before shipping
tts_text = (
    'Welcome to our weekly update. '
    'This week, the engineering team shipped three major features. '
    'First, we improved search speed by forty percent. '
    'Second, we added support for twelve new languages. '
    'Third, we launched our new mobile application.'
)
path = generate_and_test_tts(tts_text)

Выбор голоса и согласование запроса

VOICE_SPECIFIC_PROMPTS = {
    'professional': (
        'Write in a clear, neutral, professional tone. '
        'Use complete sentences. Avoid contractions. '
        'Suitable for business reports and documentation.'
    ),
    'warm_narrative': (
        'Write in a warm, engaging, conversational tone. '
        'Use contractions naturally. '
        'Speak directly to the listener using "you" and "we". '
        'Suitable for educational content and storytelling.'
    ),
    'energetic': (
        'Write in an upbeat, enthusiastic tone. '
        'Use shorter sentences for impact. '
        'Include emphasis words like "amazing", "incredible", "now". '
        'Suitable for marketing and product announcements.'
    ),
}

# Select based on use case
use_case = 'warm_narrative'
print(VOICE_SPECIFIC_PROMPTS[use_case])

Проектирование конвейера LLM–TTS

Проверка знаний: структура предложений для TTS

Итоги: шаблоны запросов TTS для естественной речи