Text-to-Speech APIs for Phone Calls: 19 Voices Tested on Latency, Numbers and Price (Nov 2026)

Results#

What we found#

1. Three voices start talking in about 0.2 seconds#

Inworld TTS-2 Flash, ElevenLabs Flash v2.5 and Rime Mist v3 all had their first audible word out in 0.18 to 0.24 s at the median and under 0.3 s at the 90th percentile, and none of them opened their clips with silence. The new voices ran about four hours after the first batch, so we ran one more interleaved round of all five sentences for the seven fastest voices together. The order inside the top three changed (Inworld TTS-2 Flash 0.19 s, Rime Mist v3 0.20 s, ElevenLabs Flash v2.5 0.25 s), and everything below them kept its place: Inworld TTS-2 0.30 s, Cartesia 0.35 s, Google Chirp 3 HD 0.53 s and Gemini 3.8 Flash Lite 0.65 s. Treat the top three as tied on speed and choose on price, voice and features.

2. First audio is not the same as first word#

Time to first audio byte is the usual latency number, but several models start their clip with silence. Google's Chirp 3 HD sent its first audio in 0.21 s, as fast as the leaders, but opened every clip with about 0.27 s of silence, so the caller hears the first word after about 0.53 s. Google's two Gemini voices and Deepgram Aura-2 also start with about 0.26 s of silence, Cartesia with about 0.14 s and Kokoro with about 0.43 s. We measured that silence in each clip (anything quieter than -40 dB) and added it, which is the "first audible speech" column.

In a voice agent, the voice is the last of three steps that run one after another: the speech-to-text model decides the caller has finished, the language model starts its reply, and the voice starts speaking. Every tenth of a second here is a tenth of a second of silence on the line.

3. Numbers and emails mostly survive the phone line#

We expected digits and email addresses to be where cheap voices and 8 kHz audio break. They mostly did not. Every model read "415-867-5309" as a phone number both judges wrote down correctly, every model read "dana.kowalski@gmail.com" as a spelled-out address, and every model got "$249.99", "March 14th" and "3:15 PM" across in all three runs. Nobody needed SSML or a rewritten prompt for this. That is good news for builders, but it is also a small test: one phone number, one email and one address, in English, with a common name.

4. Two OpenAI results to check in your own templates#

tts-1 misread a house number. In all three runs, both judges heard "1428 Elm Street" as "Fort 128" or "128 Elm Street", and a third pass on the full-quality audio heard the same. The phone line did not cause it. If you use tts-1 for addresses, write house numbers out in words or test your formatting first.

gpt-4o-mini-tts left out a sentence. For "I have your number as 415-867-5309 and your email as dana.kowalski@gmail.com. Is that right?", the audio ended after the email in all three runs. Both judges, and OpenAI's own gpt-transcribe on the last three seconds of the clip, heard the email read slowly and then nothing. The closing question is the part that turns a read-back into a confirmation, so this is the kind of omission a caller notices. gpt-4o-mini-tts is a steerable model that takes style instructions; we sent none. We saw it on one sentence of five, three times out of three.

The other small differences: Kokoro (3 of 3 runs), Inworld TTS-2 Flash and Rime Coda (1 of 3 each) said "your email is" where the text said "your email as", and one judge heard Deepgram Aura-2 say "their" for "your" once. None changes the meaning.

5. The same model through a reseller can be ten times slower#

Inworld TTS-2 through its own API started speaking in 0.28 s. The same model through DeepInfra took 2.64 s, because DeepInfra's endpoint appears to make the whole clip before sending any of it, and DeepInfra charges $35 per 1M characters against Inworld's own $25. MiniMax and xAI, which we only reached through fal.ai, return a finished file, so they only have a "whole sentence" time (3.5 to 4.6 s including fal's queue); both stream on their own APIs, which we did not test.

Price per minute, tested and untested#

Prices from each vendor's own page, read November 12, 2026. For per-character prices we convert at about 740 characters per minute of speech, the median speaking rate of the voices we tested (they ranged from 631 to 911 characters per minute, so a fast voice costs more per minute of audio than a slow one at the same per-character price). Native telephony format means the API can return 8 kHz mu-law (G.711) itself, so you do not have to convert the audio before it goes onto a call.

If you are leaving a voice that is going away, see Hume AI alternatives and PlayHT alternatives.

Worth a look

ElevenLabs

Lifelike AI voiceovers, dubbing and voice cloning in dozens of languages. Free plan to start.

Try ElevenLabs

One weakness each#

• Inworld TTS-2 Flash: no natural-language voice steering, which Inworld reserves for TTS-2 (model docs). Stork earns a commission on Inworld sign-ups, so weigh our enthusiasm against that; the numbers are measured the same way for every vendor.
• ElevenLabs Flash v2.5: the most expensive of the three fastest voices, about three times Inworld TTS-2 Flash and four times Gemini 3.8 Flash Lite per minute at list price. It also speaks faster than most voices, so the same text makes fewer minutes but each minute costs more.
• Rime Mist v3: billed per minute of audio, at $0.030 almost three times Inworld TTS-2 Flash.
• Inworld TTS-2: took 1.8 s to finish a sentence against 0.7 s for TTS-2 Flash, and through a reseller (DeepInfra) it was ten times slower to start.
• Smallest.ai Lightning v3.1 Pro: a step behind the three fastest voices (0.30 s to the first word) and about 50% more per minute than Inworld TTS-2 Flash ($0.017 against $0.011).
• Fish Audio S2.1 Pro: no mu-law output, 0.68 s to the first word and the slowest directly reached voice to finish a sentence (5.3 s at the median). We could only test the free developer tier, because our account has no paid API credit; the paid tier may queue differently.
• Cartesia Sonic 3.6: no plain pay-as-you-go; you buy a monthly plan and overage is off by default. It also opens each clip with about 0.14 s of silence.
• Rime Coda: the most expensive voice per minute in our test ($0.050) and the only one besides tts-1 with an entity miss (one misspelled surname in three runs).
• ElevenLabs v4 Turbo: the same price as Flash v2.5 but slower on both counts here: 0.47 s to the first word and 1.8 s to finish a sentence.
• Google Cloud Chirp 3 HD: about 0.27 s of silence at the start of every clip, and streaming needs Google's gRPC streaming API, not the plain REST call.
• Gemini 3.8 Flash Lite TTS: about 0.26 s of silence at the start of every clip, the price doubles on January 1, 2027, and 8 kHz mu-law is only on Google's newer Interactions API.
• Deepgram Aura-2: the slowest of the directly reached voices to finish a sentence (4.3 s at the median) and 0.72 s to the first word.
• gpt-4o-mini-tts: dropped the closing question of one sentence in all three runs, and the API has no 8 kHz or mu-law output.
• Gemini 3.8 Flash TTS: slower than Flash Lite on every measure here (0.99 s to the first word, 2.8 s per sentence) and 50% more expensive per minute.
• Kokoro-82M: 0.43 s of silence before each clip, preset voices only, no cloning, and on DeepInfra no telephony format.
• tts-1: the slowest streaming voice to start (1.25 s) and misread a house number in every run.
• xAI TTS v1: through fal it does not stream, so we could not measure its first-audio time; its own API streams over WebSocket.
• MiniMax Speech 2.8 Turbo: among the most expensive per minute in our test ($0.038) and the slowest file through fal (4.6 s).

How we tested#

• Sentences: "Thanks for calling Riverside Dental, this is Ava. How can I help you today?"; "I have your number as 415-867-5309 and your email as dana.kowalski@gmail.com. Is that right?"; "Our technician will come to 1428 Elm Street, Apartment 3B, Springfield, Oregon 97477."; "Your total is $249.99, and your next appointment is Tuesday, March 14th at 3:15 PM."; and one 215-character instruction about what to bring and how to reschedule. Plain text, no SSML, no pronunciation hints, the same for every model.
• Requests: one sentence per request, 3 rounds, models taking turns so a network slowdown hits all of them. One unscored warm-up request per model first, so connection setup is not counted. Streaming endpoints were read as they arrived; first audio = sending the request to the first audio bytes. Ten voices ran in a first batch, seven (Cartesia, Deepgram, both Rime models, both Inworld models direct, Google Cloud) about four hours later and two (Smallest.ai, Fish Audio) after that, all on November 12, 2026; the seven fastest were then re-run together for one interleaved round as a cross-check. Google Cloud was called through its gRPC streaming API, the others over HTTPS.
• Voices and formats: ElevenLabs "Adam", OpenAI "alloy", Google Gemini "Kore", Google Cloud "en-US-Chirp3-HD-Kore", Inworld "Ashley", Cartesia "Jacqueline", Deepgram "Thalia", Rime "astra", Smallest.ai "chelsea", Fish Audio the popular English community voice "Sarah" with latency set to "low", Kokoro "af_bella", fal's defaults for MiniMax and xAI. Every vendor that offers it was asked for native 8 kHz mu-law (ElevenLabs, Inworld direct, Cartesia, Deepgram, Rime, Smallest.ai, Google Cloud); Fish Audio sent 8 kHz PCM and the others 16 to 24 kHz audio.
• Phone line: every clip was band-passed to 300-3,400 Hz, resampled to 8 kHz and encoded to G.711 mu-law, then decoded, the same chain as our speech-to-text phone audio test. It simulates a phone line in software, not a real carrier path with packet loss.
• Judges: two speech-to-text models from vendors with no voice in this test, Whisper large-v3 and Cohere Transcribe (both through fal), transcribed all 285 phone-line clips. Scoring uses OpenAI's Whisper English normaliser with each digit counted as its own word, so "415-867-5309" and "four one five..." score the same; email addresses are compared spelled out. An entity counts as wrong only if both judges missed it, so one judge's mistake is not charged to the voice. The two OpenAI problems reported above were rechecked with a third transcription: Whisper large-v3 on the full-quality tts-1 audio, and OpenAI's gpt-transcribe on the last seconds of the gpt-4o-mini-tts clip.
• Leading silence: measured in each clip as the time before the first sound louder than -40 dB.
• Prices: from each vendor's own page on November 12, 2026, linked in the table. Price per minute for tested voices = list price for the characters (or, for Gemini, the audio tokens Google reported; for Rime, the minutes) divided by the minutes of audio returned. Cartesia is priced at its Startup plan rate ($49 for 1.25M characters). OpenAI does not return audio-token counts on a streamed request, so gpt-4o-mini-tts uses OpenAI's own estimate of $0.015 per minute.
• Cost of this test: about $1.70 in API calls at list prices (logged per call; Google Cloud's share is small enough to sit within its monthly free tier, and Deepgram's came from prepaid credits), plus several hundred ElevenLabs plan credits. The harness is in the Stork repository at scripts/bench/tts_phone_bench.py, and the raw audio and transcripts are kept so every number can be checked.

What this test does not tell you#

• How the voices sound. We did not score naturalness, warmth or accent. Listen to the voice you plan to use on a real phone before you choose it.
• Your location and your network. One client in California. Absolute latencies depend on where your servers are; compare models within this table.
• WebSocket streaming. We used each vendor's HTTP streaming endpoint (gRPC for Google Cloud). Voice agents often stream text in over a WebSocket as the language model writes it, which can start audio earlier.
• Sample size. 5 sentences, 3 rounds, one voice per model, English only. Differences under about 0.05 s between the fastest voices are within run-to-run variation. A model that read our one email correctly may still trip over yours.
• Untested vendors. Azure, Amazon Polly, Murf, Speechify and Deepgram's Flux TTS were not tested because we have no keys or did not include the model this round, and Fish Audio's Drama 3 because our account has no paid API credit. A missing model is not a worse model.