Choose TTS for an audiobook.

Current narration candidates

Capabilities below come from the linked vendor documentation. The old cross-vendor MOS, price and long-form radar scores lacked a shared protocol and inspectable runs; they have been removed. This page does not establish an overall naturalness winner.

Chirp 3 HD supports a specific SSML subset

Google lists speak, say-as, p, s, phoneme, sub, break, audio, prosody and voice. The say-as expletive/bleep interpretations are excluded. This is preview support for synchronous synthesis, not full SSML 1.1 or streaming support. Google SSML support matrix ↗.

The following example uses supported elements in a synchronous request. It follows the Cloud TTS client shape in the documentation and has not been executed against a billed API during this review.

# pip install google-cloud-texttospeech
# Configure Google Application Default Credentials first.
from google.cloud import texttospeech

client = texttospeech.TextToSpeechClient()
response = client.synthesize_speech(
    input=texttospeech.SynthesisInput(
        ssml='<speak><p><s>Chapter one.</s></p><break time="500ms"/>The story begins.</speak>'
    ),
    voice=texttospeech.VoiceSelectionParams(
        language_code="en-US", name="en-US-Chirp3-HD-Charon"
    ),
    audio_config=texttospeech.AudioConfig(
        audio_encoding=texttospeech.AudioEncoding.MP3
    ),
)
with open("chapter-sample.mp3", "wb") as output:
    output.write(response.audio_content)

A generic PLS lexicon file is not automatically a supported Chirp input. Use the documented pronunciation mechanisms for the selected voice. Azure HD also needs its family-specific support matrix; standard Azure Neural features cannot be assumed to work in DragonHD or Omni.

Build and review a chapter before rendering a book

1. Create a short test chapter covering narration, dialogue, recurring names, abbreviations, dates, quotations and emotional changes.
2. Pin the model, voice and supported control settings. Keep the same narrator configuration across chunks; record every render version.
3. Split at sentence and scene boundaries within the service's request limits. Review transitions for pauses, clipping and voice drift.
4. Maintain a pronunciation list and encode it using the selected model's supported controls. Validate corrections in context.
5. Listen through the whole chapter, check it against the manuscript, record missing/repeated words and rerender errors. Apply mastering settings required by the actual distributor.

Before comparing quality, use identical passages and speaker conditions with a blind listener protocol. Store audio, model IDs, generation settings and revision counts. A short clean sample does not establish ten-hour narrator stability.

Distribution and rights

Confirm the selected distributor's current AI-narration eligibility, disclosure and audio specifications before submission. Access to a pilot or a vendor voice does not imply acceptance by every marketplace. Verify narrator consent and the specific commercial terms of the voice or checkpoint.

ACX help and submission requirements ↗ · Local model code and weight licenses · Voice-agent latency evaluation.