Text to Speech (TTS) in 2026: How Neural Voices Work, Top Tools Compared, and TTS for Phone Calls

Text to speech (TTS) is technology that converts written text into spoken audio. Modern neural text to speech uses deep learning models trained on recorded speech to produce voices with natural rhythm, intonation and pronunciation, in many languages. It powers screen readers, audiobooks, video voiceovers, navigation apps and, increasingly, AI voice agents that speak on live phone calls.

This guide explains how neural TTS works, what separates a realistic AI voice from a merely clear one, the consent rules around voice cloning, and how seven major text to speech tools compare on pricing and features as of 28 October 2026. The second half covers TTS for phone calls, where narrowband audio and a tight latency budget change what good looks like, and when you need a full voice agent rather than a TTS API.

What is text to speech (TTS)?

A text to speech system takes text, works out how it should be said, and generates the audio. Written language leaves out much of what a speaker knows: whether “read” is past or present tense, how to say “Dr.” or “12/10”, where to pause, and which word in a sentence carries the stress. Good TTS resolves all of that before a sound is produced.

Older systems either stitched together fragments of recorded speech (concatenative synthesis) or generated audio from statistical acoustic models (parametric synthesis), and both sounded noticeably artificial. Neural TTS replaced them with models that learn the mapping from text to sound directly from many hours of recorded speech.

How does neural text to speech work?

Most neural TTS systems work in stages, even when they are packaged as a single model:

1. 01Text normalization: numbers, dates, currencies, abbreviations and symbols are expanded into words, so “₹1,499” becomes “one thousand four hundred ninety-nine rupees”.
2. 02Linguistic analysis: words are converted into phonemes, and the model predicts prosody, meaning the rhythm, stress and intonation of the sentence.
3. 03Acoustic modeling: a neural network predicts an intermediate representation of the sound, typically a mel spectrogram or a sequence of discrete audio tokens.
4. 04Waveform generation: a neural vocoder or audio codec decoder turns that representation into the audio samples you hear.

Three research milestones show how fast this moved. The WaveNet paper (2016), from DeepMind researchers, introduced a deep neural network that generates raw audio waveforms, and listeners rated its speech significantly more natural than the best parametric and concatenative systems of the time. The Tacotron 2 paper (2017) paired a network that predicts mel spectrograms with a WaveNet vocoder and reported a mean opinion score of 4.53, against 4.58 for professionally recorded speech. In 2023, the VALL-E paper treated TTS as language modeling over discrete codes from a neural audio codec, and could imitate an unseen speaker from a 3-second recording.

Today's commercial voices build on these ideas. Many newer models generate audio as a sequence of tokens, an approach that lends itself to streaming output and to cloning from short samples.

What makes realistic AI voices sound human?

A demo tells you only part of the story. Judge a TTS voice on six things:

• Naturalness: whether listeners hear it as human. Research uses the mean opinion score (MOS), a 1 to 5 listener rating; in practice, run a blind listening test with your own scripts.
• Prosody: rhythm, stress and intonation. A voice can be clear and still sound wrong if it stresses the wrong word or reads a question flat.
• Pronunciation and SSML: names, product terms, numbers and addresses are where voices fail. The W3C's Speech Synthesis Markup Language (SSML 1.1, a 2010 Recommendation) controls pronunciation, pauses, pitch, rate and volume with elements such as say-as, phoneme, break and prosody, but support varies. Google, for example, lists SSML for its Chirp 3: HD voices as Preview and does not support it in streaming requests.
• Latency and time to first audio: how long until the first sound plays. For a narrated video it barely matters; on a live call it decides whether the conversation feels natural.
• Languages and accents: check the exact languages, regional accents and mixing you need, such as Hindi with English, rather than a headline language count.
• Streaming: whether the API returns audio in chunks while it is still generating, so playback starts before the full sentence is ready.

Consistency matters too. A voice that sounds great for one sentence can drift in tone or pace over a five-minute call, so test long, multi-turn samples, not single lines.

Voice cloning: what are the consent and disclosure rules?

Voice cloning creates a synthetic copy of a specific person's voice, sometimes from seconds of audio; Cartesia, for example, says its instant cloning works from 10 seconds. That is why the major platforms put consent checks in front of it:

• Microsoft: before fine-tuning a professional voice in Azure, you must upload a recording of the voice talent reading a consent statement, which is also used to verify that the speaker matches the training data. Custom voice is a limited access feature.
• Google: Instant Custom Voice is restricted to allow-listed users and requires a recorded consent statement, “I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model”, with no custom wording allowed.
• ElevenLabs: every Professional Voice Clone goes through verification to confirm the voice belongs to the account holder, and its documentation says that even with consent you cannot clone someone else's voice; that person must create and share the clone from their own account.
• OpenAI: custom voices are created from a speaker's consent recording, and OpenAI's usage policies require a clear disclosure to end users that the TTS voice they hear is AI-generated and not a human voice.

Law is catching up. In the US, the FCC's February 2024 declaratory ruling confirmed that AI-generated voices, including voice cloning, count as an artificial voice under the Telephone Consumer Protection Act, so calls using them need prior express consent. In the EU, the AI Act's Article 50 transparency duties have applied since 2 August 2026: people must be told when they are interacting with an AI system unless it is obvious, and providers of systems that generate synthetic audio must mark it in a machine-readable way. A Baker Botts summary notes a grace period to 2 December 2026 for marking by systems already on the market before August.

Text to speech tools compared (checked 28 October 2026)

The table lists seven widely used TTS tools, using only pricing and features stated on each vendor's own pages on 28 October 2026. As a rough conversion, Amazon's pricing example equates 1 million characters with about 23 hours of speech, and Cartesia estimates one minute of audio at 750 to 800 characters.

How do you choose among text to speech tools?

1. 01Start from the use case. Narration, e-learning and voiceovers reward the most expressive voice; live conversation rewards streaming and fast first audio.
2. 02Test with your own scripts, including names, numbers, prices, addresses and the languages your audience speaks.
3. 03Check output formats. Traditional phone lines carry 8 kHz μ-law or A-law audio, while video and podcasts call for high-quality formats such as WAV or MP3.
4. 04Read the cloning and commercial-use terms, and keep consent records for any custom voice.
5. 05Compare cost per finished minute, not per character, at your real monthly volume.

TTS for phone calls: what changes on a telephone line?

A voice that sounds rich through headphones can sound thin on a call, because traditional telephone audio is narrowband. ITU-T Recommendation G.711, the classic telephony codec, specifies a nominal sampling rate of 8,000 samples per second, which limits audio to frequencies below 4 kHz. Streaming APIs follow the same format: Twilio's Media Streams, for example, sends and expects mono μ-law audio at 8,000 Hz.

• Request telephony-ready audio. Generating μ-law or A-law at 8 kHz directly avoids a resampling step; Google's streaming synthesis, for example, can return MULAW and ALAW output.
• Test voices on a real phone line. Breathy or very soft voices tend to lose detail in narrowband audio, while clear, steady voices hold up better.
• Write for the ear. Short sentences, one question at a time and numbers read in natural groups are easier to follow on a call than long paragraphs.
• Check pronunciations for names, addresses, order numbers and currencies in each language you call in.

How fast does TTS need to be for a live conversation?

People notice silence quickly. A 2009 PNAS study of turn-taking in 10 languages found that speakers everywhere avoid overlapping talk and minimize silence between turns, with average gaps differing across languages only within a range of 250 milliseconds of the overall mean. A voice agent has to fit its reply into that rhythm.

TTS is only one part of that budget. A full turn includes detecting that the caller has finished, transcribing the words, generating a reply with a language model, any data lookup, synthesizing the first audio and carrying it across the phone network. Streaming makes it work: the TTS starts speaking the first phrase while the rest of the reply is still being generated. Vendors advertise first-audio figures of about 75 to 100 ms (ElevenLabs), under 90 ms (Cartesia) and as low as 80 ms (Deepgram), each measured under their own conditions.

What the caller feels is the end-to-end number, not the TTS number. Vaami's voice agents deliver sub-300ms end-to-end response latency in production and handle interruptions (barge-in), so callers can talk over the agent the way they would with a person.

When do you need a full voice agent instead of a TTS tool?

A TTS API turns text into audio. It does not answer the phone, understand the caller, decide what to say, look up a booking or hand over to a colleague. For narration, e-learning, IVR prompts or voiceovers, a TTS tool is exactly right. For phone conversations, TTS is one component of a larger system that also needs:

• Telephony: numbers, SIP trunks, call control and transfers.
• Speech recognition tuned for phone audio, accents and code-switching.
• A language model with guardrails, your knowledge base and business rules.
• Turn-taking: detecting when the caller has finished, handling interruptions, and filling pauses during lookups.
• Actions: CRM updates, calendar booking, SMS or WhatsApp follow-ups and webhooks.
• Analytics and compliance: transcripts, QA scoring, consent capture, disclosure and calling-hour rules.

Vaami is not a TTS tool. It uses TTS inside AI voice agents for inbound and outbound calls, in 30+ languages including Hindi, Hinglish, Tamil, Telugu, Arabic and Spanish, with warm transfers to your team and native connectors for 100+ platforms. Standard configurations go live in under one business day, and AI usage costs $0.025 per minute, billed per second, with telephony separate. The quickest way to judge the voice is to hear it on a real call.

Sources

• van den Oord et al., WaveNet: A Generative Model for Raw Audio, arXiv, 12 September 2016: arxiv.org/abs/1609.03499
• Shen et al., Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Tacotron 2), arXiv, 16 December 2017: arxiv.org/abs/1712.05884
• Wang et al., Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E), arXiv, 5 January 2023: arxiv.org/abs/2301.02111
• W3C, Speech Synthesis Markup Language (SSML) Version 1.1, Recommendation, 7 September 2010: w3.org/TR/speech-synthesis11
• Google Cloud, Text-to-Speech pricing, accessed 28 October 2026: cloud.google.com/text-to-speech/pricing
• Google Cloud, Chirp 3: HD voices, accessed 28 October 2026: docs.cloud.google.com/text-to-speech/docs/chirp3-hd
• Google Cloud, Chirp 3: Instant custom voice, accessed 28 October 2026: docs.cloud.google.com/text-to-speech/docs/chirp3-instant-custom-voice
• Amazon Web Services, Amazon Polly pricing, accessed 28 October 2026: aws.amazon.com/polly/pricing
• Microsoft, Azure Retail Prices API, Azure Speech meters for East US, accessed 28 October 2026: prices.azure.com/api/retail/prices
• Microsoft Azure, Speech services pricing, accessed 28 October 2026: azure.microsoft.com/en-us/pricing/details/cognitive-services/speech-services
• Microsoft Learn, Add voice talent consent for professional voice, updated September 2026: learn.microsoft.com/en-us/azure/ai-services/speech-service/professional-voice-create-consent
• ElevenLabs, API pricing, accessed 28 October 2026: elevenlabs.io/pricing/api
• ElevenLabs Docs, Professional Voice Cloning, accessed 28 October 2026: elevenlabs.io/docs/product-guides/voices/voice-cloning/professional-voice-cloning
• OpenAI, Text to speech guide, accessed 28 October 2026: developers.openai.com/api/docs/guides/text-to-speech
• Cartesia, Pricing, accessed 28 October 2026: cartesia.ai/pricing
• Cartesia, Sonic, accessed 28 October 2026: cartesia.ai/sonic
• Deepgram, Pricing, accessed 28 October 2026: deepgram.com/pricing
• Deepgram, Text-to-speech, accessed 28 October 2026: deepgram.com/product/text-to-speech
• ITU-T, Recommendation G.711, Pulse code modulation (PCM) of voice frequencies, November 1988: itu.int/rec/T-REC-G.711
• Twilio Docs, Media Streams WebSocket messages, accessed 28 October 2026: twilio.com/docs/voice/media-streams/websocket-messages
• Stivers et al., Universals and cultural variation in turn-taking in conversation, PNAS, 30 June 2009: pnas.org/doi/10.1073/pnas.0903616106
• Federal Communications Commission, Declaratory Ruling FCC 24-17 on AI technologies and robocalls, adopted 2 February 2024: docs.fcc.gov/public/attachments/FCC-24-17A1.pdf
• Baker Botts, EU AI Act Article 50 transparency obligations go live, 10 September 2026: bakerbotts.com/thought-leadership/publications/2026/september/eu-ai-act-article-50-transparency-obligations-go-live