Text-to-speech narration (audiobooks, e-learning, IVR)

Recording narration in a studio is slow and expensive, and every text change means a re-record. Neural text-to-speech (TTS) now produces near-human, expressive voices in dozens of languages from a script, in minutes and at cents per minute of audio. It is used for audiobooks, course narration, phone menus, news readers, accessibility and, increasingly, AI-generated audio briefings and podcasts from documents. Since AI Act Art. 50 (applicable since 2 Aug 2026; Art. 50(2) marking for generative systems already on the market: 2 Dec 2026), synthetic audio must be marked as machine-generated.

Problem & value

• Today: audio is produced with voice actors and studios, or not at all. Publishers leave most backlist titles unrecorded; training teams skip narration or never update it; IVR prompts are re-recorded for every change and every language; many websites offer no audio version.
• AI value: audio on demand and in many languages, instant updates (change one sentence, re-render one clip), consistent voice, lower cost per finished minute, and accessibility for people with visual impairments, dyslexia or low literacy. Human narrators keep their value for premium, emotionally demanding titles.
• Market evidence: large audiobook platforms now accept or offer AI-narrated titles with disclosure (e.g. Google Play Books auto-narration, Audible's AI narration programme), and Google NotebookLM popularised multi-speaker "Audio Overviews" generated from user documents [3][4]. Do not rely on vendor claims about savings; measure on your own content.

Variants

• Audiobooks & publishing: long-form, single or multi-voice narration of EPUB/manuscripts with chapter handling, pronunciation lexicons and human proofing; often hybrid (AI first pass, human QA or human lead voice).
• E-learning & corporate training: slide/script narration, synchronised captions, instant re-render when content changes, 20+ languages from one script (see EDU-006).
• IVR & telephony prompts: static prompt libraries and dynamic prompts (names, amounts, dates) in 8 kHz telephony quality; low-latency streaming TTS when part of a voicebot (CONV-002, CONV-008).
• AI audio briefings & podcasts from documents (format variant): an LLM turns reports, papers or meeting packs into a two-host dialogue script (summarisation, see LANG-008) and multi-speaker expressive TTS renders it. Evidence: NotebookLM Audio Overviews [4], and similar features in ElevenLabs and OpenAI offerings. It is a combination of existing building blocks, not a new capability; the main risk is faithfulness of the generated script to the source.

How it works

1. source text (CMS, EPUB, DOCX, LLM-generated script)
2. text normalisation (numbers, dates, abbreviations, language tagging)
3. optional LLM pass: script rewrite for listening, dialogue script, SSML/emotion tags
4. pronunciation lexicon (names, product terms, IPA/phoneme overrides)
5. TTS model (voice, language, style, speed)
6. audio + word timestamps
7. post-processing (loudness normalisation, silence trimming, music bed, format/codec)
8. human review (spot-check, proof-listen of critical passages)
9. marking/disclosure (metadata watermark, "AI voice" label)
10. publish / IVR deploy

• Model families: autoregressive codec-language-model TTS and diffusion/flow-matching TTS generate speech from text plus a voice prompt; streaming variants deliver the first audio in well under a second for voicebots.
• Controllability: SSML tags, inline emotion/style tags, and lexicons; long-form quality depends on chunking with consistent voice and prosody across chapters.
• Human-in-the-loop: proof-listening of names, numbers and legal text; approval of voice and rights; sign-off before public release. Any safety-relevant prompt (e.g. IVR for emergencies, medication instructions) must be human-verified.

Data preparation

TTS needs little training data: the source text (manuscripts, e-learning scripts, IVR prompts, articles), a pronunciation lexicon and a decision on the voice. In an SMB the texts live in Word or PDF files on file shares, in a CMS or LMS, or in the telephony system as old recordings.

Typical gaps: scanned or layout-heavy PDFs that do not convert cleanly, inconsistent product names and acronyms, numbers and abbreviations that are read wrongly, duplicated script versions and unclear rights to texts and voices.

Preparation steps: 1. Convert scans and PDFs to clean text; use OCR where needed (PREP-006, PREP-007). 2. Normalise text: expand abbreviations, spell out numbers, mark language switches, add SSML pauses. 3. Build a pronunciation lexicon for brand names, acronyms and place names (PREP-015) and keep it in one shared file. 4. Pick one approved voice, record the licence or consent, and create a short reference script as test set (PREP-016, PREP-018). 5. Version scripts in one place so updated prompts replace old audio.

Rough share of total effort: about 10 to 15 percent; most time goes into text cleanup and listening tests. See GUIDE-003 for a quick check.

Implementation paths

Which path for whom

• smb: Path A, because mature EU-hostable TTS products with licensed voices and simple web editors exist at SMB prices and narration is a standard process.
• enterprise: Path B, because large IVR and e-learning libraries are best generated through an API in existing content and telephony workflows, with volume pricing.
• developer: Path B, because TTS APIs with streaming and SSML integrate directly into own apps with little effort.
• public: Path C, because authorities favour sovereign or on-premises operation and open-source voices avoid procurement and licensing hurdles for public announcements.
• private: Path C, because local apps and open-source voices (Piper, Kokoro-class) run on a normal laptop and keep texts private.

Path A — Buy (recommended for most SMBs)

• Voice-over and narration studios (SaaS): ElevenLabs (also Reader, audiobook publishing), Murf, WellSaid, Speechify, LOVO. EU: ReadSpeaker (Sweden), Acapela Group (France/Belgium), further regional vendors; many e-learning tools (Articulate 360, Synthesia, Camtasia, Microsoft Clipchamp) embed TTS.
• Audiobook platforms: Google Play Books auto-narration, Audible (AI narration programme), subject to current publisher terms.
• IVR/contact centre: TTS is built into the telephony platform (Genesys, Amazon Connect, Twilio, Telekom/Vodafone business platforms, 3CX).
• Choose A when: audio volume is moderate, you need a polished editor, and licensing of stock voices is acceptable. Check EU data processing, retention of uploaded text, voice licence terms and commercial-use rights.

Path B — Build on cloud APIs

• Hyperscaler TTS: Azure AI Speech (neural and HD voices, custom neural voice, EU regions), Google Cloud Text-to-Speech (Chirp 3 HD, Gemini-TTS, EU locations), Amazon Polly (Frankfurt/Ireland regions, neural and generative voices) [1][2].
• Specialist APIs: ElevenLabs API (v3 expressive model, Flash for low latency, EU data residency options for enterprise), OpenAI TTS API (gpt-4o-mini-tts with instructable style), Cartesia, Deepgram Aura. EU/UK: Speechmatics (UK), plus European vendors above.
• Choose B when: you integrate narration into your own app, LMS or CMS.

Path C — Self-host (open source)

• Open-weight models (check each licence for commercial use): Kokoro-82M (Apache-2.0, tiny, fast), Piper (MIT/GPL variants, runs on CPU and Raspberry Pi, many EU languages), Coqui XTTS-v2 (non-commercial CPL licence, use with care), Fish Speech / OpenAudio, Chatterbox (MIT, with built-in watermarking), Orpheus-TTS, Dia, F5-TTS, Parler-TTS [5][6]. German and other EU-language quality varies strongly between models; test with your own text.
• Hardware: Piper and Kokoro run on CPU/HW1; LLM-based expressive models need one 8-24 GB GPU (HW1-HW2) for real-time or faster-than-real-time synthesis.
• Choose C when: text is confidential (unreleased manuscripts, internal HR content), volume is large, or sovereignty requires on-prem operation.

Path D — Custom training

Worth it for a brand or narrator voice (with the voice owner's explicit written consent), a rare language or dialect, or domain vocabulary. Typically a few hours to tens of hours of clean, consistently recorded studio audio with transcripts; commercial services also offer professional voice clones from 30 minutes to a few hours. Strongly consider consent, contract and revocation terms first (see SPEECH-007).

State of the art (2026)

• Quality: top models are often rated close to human speech in short passages in listening tests; community arenas such as TTS Arena (Hugging Face) rank open and commercial models by blind preference. Expressive tags, emotion control and multi-speaker dialogue are standard in leading products.
• Open models closed much of the gap: Kokoro, Chatterbox, Fish Speech and similar models give good English quality on commodity hardware; multilingual EU coverage and prosody consistency over hours still lag the best commercial systems.
• Failure modes: mispronounced names, numbers, homographs and abbreviations; drift in voice or pacing across long chapters; hallucinated or skipped words in autoregressive models; flat prosody for dialogue and poetry; poor 8 kHz telephony rendering if tuned for studio audio; dialect gaps.
• Standards and provenance: C2PA content credentials and audio watermarking (e.g. Google SynthID, vendor watermarks) are used to mark synthetic audio, but no watermark is robust to every re-encoding; combine metadata, visible labels and logs.

Shared resources & combinations

• Same capability (tts): SPEECH-007 voice cloning & dubbing, SPEECH-014 audio description, SPEECH-018 assistive speech, SPEECH-019 voice conversion, CONV-002 / CONV-008 voice agents, CONV-014 in-car assistants.
• Upstream: LANG-008 (summaries to narrate), LANG-translation workflows and SPEECH-003 for multilingual versions, DOC-026 (accessible documents give clean reading order for TTS), EDU-006 (training content).
• Downstream: SPEECH-004 / SPEECH-008 (captions, loudness and mastering), podcasts and video tools.
• Hardware: one GPU (HW2) can serve TTS alongside ASR and a small LLM for a complete voice stack.

Effort, cost & KPIs

• Effort: Path A days; B 1-4 weeks; C 2-6 weeks; D months. Cost scales with characters synthesised; cache repeated IVR prompts.
• KPIs: cost per finished audio minute, turnaround time, listener ratings (MOS or A/B vs. human), pronunciation error rate per hour, listen-through rate and completion, IVR containment and caller satisfaction, share of content with audio version, share of audio correctly labelled.

People & change

Involve the department lead who owns the content (training, marketing, customer service), end users who listen to or maintain the audio, IT for telephony or LMS integration, and legal for voice licences and rights. Management decides on voice and brand tone.

Works council: usually no, because the output is audio for customers or learners and does not monitor employees. It depends only if the generated narration replaces employee speakers or is used in HR training with learner tracking; then §87(1) no. 6 and §90(1) no. 3 BetrVG apply. Inform the council early anyway if staff voices are replaced.

Data protection: mostly no personal data in scripts; check texts with customer data and use an AVV (Art. 28 GDPR) with the TTS provider. Involve the DPO only if personal data is read out.

Transparency: Art. 50 AI Act requires that synthetic audio is disclosed where it could be taken for a human, and marking applies to generators (see GUIDE-014).

AI literacy: a short workshop on proofreading, pronunciation checks and voice licensing (GUIDE-006).

Pilot: start with one e-learning module or one IVR menu. Typical concern: loss of work for voice talent and quality of the voice. Keep a human listening check before release and, for audiobooks, label synthetic voices openly. See GUIDE-008 and GUIDE-009.

Risks, regulation & compliance

• EU AI Act: limited risk. Art. 50(2) requires providers of AI systems generating synthetic audio to mark outputs in a machine-readable format; Art. 50(4) requires deployers of deepfake audio (voices imitating real persons that could be mistaken for authentic) to disclose it. For clearly artistic or creative works the duty is limited to disclosure that does not hamper enjoyment; text published to inform the public is exempt only under human editorial control. Obligations apply since 2 Aug 2026 (Art. 50(2) marking for generative systems placed on the market before that date: until 2 Dec 2026, Art. 111(4), per the Digital Omnibus on AI); follow the Commission code of practice on marking and labelling [7][8]. Telephone bots must also tell callers they are speaking to AI (Art. 50(1)). Art. 4 AI literacy applies. Using a cloned voice of a real person without consent is a legal risk even where not prohibited (see SPEECH-007).
• GDPR: a voice is personal data; a voice clone of a real narrator or employee needs a legal basis (consent plus contract), purpose limitation, deletion on request. Texts may contain personal data (customer names in IVR, HR content); sign a DPA, choose EU regions, check whether the vendor trains on your inputs, and avoid sending sensitive data to US services. Voice data used for biometric identification would be special-category data (not the case here).
• Copyright and personality rights: rights to text, to the narrator's voice (contracts for voice actors with AI clauses, collective agreements), and to training data of the TTS model; check commercial-use licences of open models (e.g. XTTS non-commercial). Platform rules may require disclosure of AI narration.
• Quality and safety: wrongly pronounced numbers, drug doses or legal terms can mislead; keep human verification for safety- and legally relevant audio.

Getting started

1. Pick one pilot (e.g. 20 help articles, one course module or the IVR menu) and write 10 hard test sentences with names, numbers and abbreviations.
2. Compare 3 voices/models blind with 5 listeners: one SaaS (Path A), one cloud API with EU region (Path B), one open model (Path C).
3. Build a pronunciation lexicon and an SSML or style template; set up loudness normalisation.
4. Add human proof-listening and an "AI-generated voice" label plus metadata marking.
5. Clear voice licences and your data-processing terms before publishing; measure KPIs for a month.

Sources

1. Microsoft — Azure AI Speech text to speech — https://learn.microsoft.com/azure/ai-services/speech-service/text-to-speech
2. Google Cloud — Text-to-Speech documentation — https://cloud.google.com/text-to-speech/docs
3. Amazon Polly — developer guide — https://docs.aws.amazon.com/polly/latest/dg/what-is.html
4. Google — NotebookLM Audio Overviews — https://support.google.com/notebooklm/answer/16212820
5. Kokoro-82M model card — https://huggingface.co/hexgrad/Kokoro-82M
6. Piper TTS project — https://github.com/OHF-Voice/piper1-gpl
7. Regulation (EU) 2024/1689 (AI Act), Art. 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. European Commission — AI Act regulatory framework — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
9. ElevenLabs — text to speech documentation — https://elevenlabs.io/docs/capabilities/text-to-speech
10. OpenAI — Text to speech guide — https://platform.openai.com/docs/guides/text-to-speech