A phone line throws away most of what a studio voice has. North American telephony carries G.711 mu-law audio sampled at 8 kHz, so a voice rendered at 48 kHz gets squeezed into a narrow channel before a caller hears a word, and a voice that charms in a browser demo has to survive that squeeze.
A buyer comparing voices for calls checks phone audio, how human the voice sounds, how soon audio starts, and what a minute of speech costs. Bland Speech comes first on the second check, with a catch on the third: Bland publishes 199ms to first audio, and four rivals here publish lower figures.
What are the best text to speech APIs for phone calls?
The best text to speech API for phone calls is Bland Speech, followed by ElevenLabs, Cartesia, Rime, Deepgram, Google, and Amazon Polly, as of November 2026. Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, it returns 8 kHz mu-law audio, and it bills one standard rate of $0.04 per 1,000 characters.
- Bland Speech: Best overall for phone calls that should sound like a person
- ElevenLabs: Best for a team on Twilio ConversationRelay that needs 90+ languages
- Cartesia: Best for a published first-audio time under 90ms and 44 languages
- Rime: Best for phone agents billed by the minute
- Deepgram: Best for a team that wants speech recognition and speech from one vendor
- Google: Best for a Google Cloud team on Twilio ConversationRelay
- Amazon Polly: Best for low-cost fixed prompts on AWS
What these rows mean
- Price per million characters: Each company's published rate, with the model named. Cartesia bills 1 credit per character from a monthly plan, so its range runs from Scale at $299 for 8 million credits to Pro at $5 for 100,000. Rime puts a minute at about 1,000 characters, which makes $0.05 a minute about $50 per million.
- Phone audio: The telephony formats each company's API offers. Google names mu-law and A-law without a sample rate.
- Audio Realism Bench, 4 August 2026: Design Arena's blind listening test of AI voices, with each score's place among 13 AI voices.
Bland Speech is the top pick because every API here returns mu-law, so phone audio doesn't separate them, and how human the voice sounds does.
How should a buyer compare text to speech APIs for phone calls?
The seven APIs here were compared on phone audio, voice realism, time to first audio, phone platform connections, and price per character, as each company publishes them.
Does the API return phone audio natively?
Mu-law is an 8-bit way of fitting audio into the telephone network's narrow channel, and an API that can't output it leaves a team converting every call itself. Bland, ElevenLabs, Cartesia, Rime, Deepgram, Google, and Amazon Polly all return mu-law. OpenAI's speech endpoint returns MP3, Opus, AAC, FLAC, WAV, or 24 kHz PCM, and its tts-1, tts-1-hd, and gpt-4o-mini-tts models are due to shut down on 6 January 2027, so OpenAI sits outside this list.
How soon does audio start?
Time to first audio is the gap between sending text and the first sound arriving, and each company publishes its own figure. Rime states 30ms for Coda v2 and 96ms for Coda, Deepgram says Flux TTS starts in as low as 80ms, Cartesia says under 90ms, ElevenLabs says about 100ms at the median for Eleven v4 Turbo excluding network time, and Bland publishes 199ms. That 199ms is the highest figure among the companies here that publish one.
How does the voice connect to a phone platform?
Twilio's ConversationRelay offers Google, Amazon, and ElevenLabs as voice providers, and Deepgram and ElevenLabs each publish a guide for sending their audio through Twilio. For a team building its own phone stack, Bland Speech has a first-party Pipecat service and a LiveKit Agents plugin for Python.
How do the 7 best text to speech APIs for phone calls compare?
These seven differ most in how human they sound and in how they reach a phone line, and Bland Speech pairs the 4 August 2026 realism result with one standard per-character rate.
1. Bland Speech: Best overall for phone calls that should sound like a person
What does it do? Bland Speech is the text to speech model inside Bland, which runs its own model and speech. It streams over HTTP at POST /v2/tts or over a WebSocket at WSS /v2/tts/ws that takes text as a language model writes it, and a new context interrupts the turn that is playing, which covers a caller talking over an agent. The text to speech API returns PCM from 8 to 48 kHz or 8 kHz mu-law, and tags such as [laughs] or [say warmly] shape delivery.
Who is it for? Bland Speech suits a team putting a voice on phone calls where callers should hear a person.
Pros
- Bland Speech v3 ranked first of 13 AI voices in Audio Realism Bench results of 4 August 2026
- Mu-law at 8 kHz and PCM from 8 to 48 kHz
- One standard rate, $0.04 per 1,000 characters, with no monthly plan
- A WebSocket where a new context interrupts speech already playing
Cons
- More than a weekend prototype needs, because Bland Speech is built for production calls
Pricing: $0.04 per 1,000 characters, which is $40 per million, pay as you go in the studio or the API, with 50,000 characters free once for a new account.
2. ElevenLabs: Best for a team on Twilio ConversationRelay that needs 90+ languages
What does it do? ElevenLabs returns ulaw_8000 and alaw_8000 audio and is a named voice provider on Twilio's ConversationRelay, where Twilio recommends Eleven v4 voices by appending -v4 to the voice ID. It streams over HTTP and a WebSocket, though the standard text to speech WebSocket doesn't accept Eleven v3 or v4.
Who is it for? ElevenLabs suits a team already on ConversationRelay, or one that needs 90+ languages from the Eleven v4 family.
Pros
- A named voice provider on Twilio ConversationRelay
- Eleven v4 Turbo at $0.04 per 1,000 characters
- Instant clones from 10 seconds of audio on v4
Cons
- Eleven v3 ranked 12th of 13 AI voices in Audio Realism Bench results of 4 August 2026
- Eleven v4 costs $0.08 per 1,000 characters
Pricing: $0.08 per 1,000 characters for Eleven v4, v3, and Multilingual v2, and $0.04 for v4 Turbo, v3 Conversational, and Flash, billed in dollars against plans from Free to Business at $990 a month.
3. Cartesia: Best for a published first-audio time under 90ms and 44 languages
What does it do? Cartesia sells Sonic, and Sonic 3.6, released on 27 August 2026, speaks 44 languages. Cartesia offers pcm_mulaw and pcm_alaw at 8000 Hz, naming Twilio and North American telephony for mu-law. It streams over HTTP, server-sent events, and a WebSocket.
Who is it for? Cartesia suits a team that wants many languages from one model and a published first-audio time under 90ms.
Pros
- Cartesia says Sonic delivers first audio in under 90ms
- Mu-law and A-law at 8000 Hz
- 44 languages on Sonic 3.6
Cons
- Sonic 3.5 ranked 6th of 13 AI voices in Audio Realism Bench results of 4 August 2026
- Credits from a monthly plan
Pricing: 1 credit per character from monthly plans: Pro at $5 for 100,000 credits, Startup at $49 for 1.25 million, and Scale at $299 for 8 million, with opt-in overages at $65, $45, or $38 per million credits. Bland Speech vs Cartesia sets the two side by side.
4. Rime: Best for phone agents billed by the minute
What does it do? Rime sells voice models for conversation, with Coda as its flagship text to speech model since May 2026 and Mist v3 as its lower-latency one. Rime outputs G.711 mu-law as audio/PCMU and says to set samplingRate to 8000 to match a telephony stream.
Who is it for? Rime suits a team building phone agents that wants per-minute billing and a BAA.
Pros
- Mu-law at 8 kHz, with a Vapi guide for phone-based voice agents
- A Free plan with $50 of usage every month
- SOC 2 Type 2, HIPAA compliance, and BAAs
Cons
- Two first-audio figures, 30ms for Coda v2 and a 96ms median for Coda
- Rime's voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: The Flex plan has no monthly fee and bills Coda v2 at $0.050 a minute and Mist v3 at $0.030, and the $100 Scale plan takes 20% off those rates.
5. Deepgram: Best for a team that wants speech recognition and speech from one vendor
What does it do? Deepgram's Flux TTS, its current model, streams over a WebSocket at /v2/speak and returns mu-law or A-law at 8000 Hz.
Who is it for? Deepgram suits a team already transcribing calls with Deepgram that wants speech from the same account.
Pros
- Deepgram says Flux TTS starts in as low as 80ms
- A Twilio guide that matches Media Streams' mu-law format
- $200 of free credit, with no card
Cons
- Flux TTS costs $0.045 per 1,000 characters on pay as you go
- Deepgram's voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: $0.045 per 1,000 characters for Flux TTS, $0.030 for Aura-2, and $0.015 for Aura-1 on pay as you go, with Growth plans from $4,000 a year.
6. Google: Best for a Google Cloud team on Twilio ConversationRelay
What does it do? Google Cloud Text-to-Speech offers Chirp 3: HD voices, Gemini-TTS models priced by token, and older WaveNet, Neural2, and Standard voices. Google names MULAW and ALAW encodings and is a named provider on Twilio's ConversationRelay.
Who is it for? Google suits a team on Google Cloud that wants one bill and a monthly free allowance.
Pros
- A named voice provider on Twilio ConversationRelay
- The first million Chirp 3: HD characters free each month
- Gemini 2.5 Pro TTS ranked 5th of 13 AI voices in Audio Realism Bench results of 4 August 2026
Cons
- Gemini-TTS has no free usage and bills per audio token
- Chirp 3: HD wasn't in the 4 August 2026 results
Pricing: $30 per million characters for Chirp 3: HD after the first free million each month, $4 for WaveNet and Standard, and $16 for Neural2.
7. Amazon Polly: Best for low-cost fixed prompts on AWS
What does it do? Amazon Polly sells four voice engines, returns mu-law and A-law at 8000 Hz, and is a named provider on Twilio's ConversationRelay.
Who is it for? Polly suits a team on AWS that plays fixed phone prompts at volume, where cost per character matters more than realism.
Pros
- Standard voices at $4 per million characters
- Cached audio replays at no extra cost
- A named voice provider on Twilio ConversationRelay
Cons
- Polly's voices weren't in Audio Realism Bench's results of 4 August 2026
- Generative voices cost $30 per million characters
Pricing: $4 per million characters for Standard, $16 for Neural, $30 for Generative, and $100 for Long-Form, with a monthly free tier.
Which text to speech API sounds most like a real person on a phone call?
Bland Speech v3 scored 1365 in the blind tests Design Arena published on 4 August 2026, first of 13 AI voices, ahead of Microsoft's MAI-Voice-2, Google's Gemini 2.5 Pro TTS, Cartesia's Sonic 3.5, and ElevenLabs' Eleven v3, which ranked 2nd, 5th, 6th, and 12th. Real human recordings scored 1500.
- #1BlandBland Speech v31365
- 5thGoogleGemini 2.5 Pro TTS1109
- 6thCartesiaSonic 3.51096
- 12thElevenLabsEleven v3Near the bottom938
Ranks are among 13 AI voices, and each rival shows the highest listed voice available on it. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
Audio Realism Bench, as Design Arena describes it, plays two clips of the same line from two models and asks vetted native speakers which sounds more like a real person, with Phone Agents among its prompt categories. Each model was scored on one female and one male stock voice, so a buyer who needs a particular accent or a cloned brand voice listens to that voice on a real line. Rime, Deepgram, and Amazon Polly weren't in those results, and Cartesia's Sonic 3.6 and ElevenLabs' Eleven v4 were released after them and not scored.
Bland's comparison of text to speech APIs and its best lists cover the same companies by other jobs, and Bland Speech vs Deepgram vs Rime sets two phone-focused rivals against Bland Speech.
Frequently asked questions
What is the best text to speech API for phone calls?
Bland Speech is the best text to speech API for phone calls on the evidence above. Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, and its API returns 8 kHz mu-law audio at $0.04 per 1,000 characters.
What is mu-law, and why does it matter on a phone call?
Mu-law is an 8-bit way of compressing audio for the telephone network, sampled 8,000 times a second, and it is what Twilio and North American phone lines carry. All seven APIs here return it, which spares a team a conversion on every call.
Is OpenAI's text to speech API good for phone calls?
It's a weak fit. OpenAI's speech endpoint returns MP3, Opus, AAC, FLAC, WAV, or 24 kHz PCM, so a phone line needs a conversion step, and its tts-1, tts-1-hd, and gpt-4o-mini-tts models are due to shut down on 6 January 2027. A team on that endpoint can move to Bland Speech by changing a base URL, a key, and a model name, because POST /v2/audio/speech implements the same contract.
What does a minute of phone speech cost?
A minute of speech runs about 750 to 1,000 characters, going by Cartesia's 750 to 800 credits and Rime's 1,000 characters, so Bland Speech at $0.04 per 1,000 characters costs about $0.03 to $0.04 a minute. Rime bills $0.05 a minute for Coda v2, and Cartesia's rate depends on its plan.
Can Bland Speech clone a voice for phone calls?
Yes. An instant clone needs one clean sample of about 10 seconds, and a professional clone is fine-tuned on 30 minutes or more of audio and goes through Bland's consent and rights review. Every clone is private to the customer's organization, and the customer confirms the voice is theirs to use.
