Miso One reads OTP codes digit by digit, spells out order IDs and email addresses, and handles US-style phone numbers correctly. Hear the real audio from those tests on this page.
Tested scripts
A curious explainer.
Kosten: 1 Credits · Anmelden, um Willkommens-Credits zu nutzen
Each clip was produced with Miso Voice 2.0 on the same request path as the studio above. Transcripts were checked word by word in September 2026.
Context: Poison-control style 1-800 numbers are a common IVR and callback script. Grouping has to stay American English, not a raw digit dump or a cardinal number.
TranscriptCall Poison Control at 1-800-222-1222 right away.
How it reads
Context: Support callbacks mix letters, digits, and punctuation. The code has to stay recoverable. 93 is spoken as a number, which we mark in the open.
TranscriptYour order A7X-93K2-QF has shipped to jane.doe94@example.com.
How it reads
Context: Short numeric codes and menu prompts are the scripts a voice agent repeats all day. A six-digit OTP has to be digit by digit, including zero.
TranscriptYour verification code is 482907. It expires in ten minutes. Press 1 for billing. Press 2 for technical support.
How it reads
Generated with Miso Voice 2.0, September 2026. Transcripts verified word by word.
Voice-agent scripts fail in specific ways. These are the limits we measured, not a claim of Gradium-level coverage.
International numbers are not reliably grouped. A UK number such as +44 20 7946 0958 can be read as a cardinal number, for example twenty-seven thousand. Do not rely on Miso Voice for digit-by-digit reading of international numbers.
Miso Voice does not support SSML tags, phoneme overrides, pause marks, or speed control. What you type is what it reads, with built-in text normalization.
Miso One is the TTS layer, not a full voice-agent platform. Bring your own telephony, speech-to-text, and dialog stack. Use Miso One to generate the audio your agent speaks.
In practice. Callers need Press 1 / Press 2 lines that stay stable across thousands of repeats. A misread digit sends them to the wrong queue.
Our approach. Pre-generate the menu lines as complete MP3 clips. The IVR clip on this page reads Press 1 for billing and Press 2 for technical support as written.
In practice. A callback that recites an order ID and a shipping email only helps if the listener can type the string back without guessing.
Our approach. Type the order ID and email as the caller should hear them. In our tests, mixed codes spelled out by default, with 93 spoken as ninety-three, and @ / . expanded in the email.
In practice. A six-digit verification code that collapses into a large number is unusable. Zero has to stay zero, not oh.
Our approach. Short digit strings such as 482907 were read one by one in our September 2026 checks, and 0 was spoken as zero. Generate the clip, then play it in the notification or bot turn.
Numbers below come from the September 2026 hard-case runs on scripts under 120 characters. They describe time to a finished clip, not a live conversational turn.
Measured September 2026 on scripts under 120 characters: average 2.65 s to a finished clip (range 1.46-4.30 s). Audio returns as a complete MP3, not a stream.
Output: MP3, 44.1 kHz, 128 kbps.
Free plan: 120 characters per render, 15 welcome credits on sign-up. Paid: 1,000 characters per render.
No SSML, phoneme overrides, or speed control.
These ranges describe human perception of a live turn. They are not a product promise.
Miso Voice 2.0 currently returns a finished clip, so it fits pre-generated prompts, IVR menus, and notification audio rather than live conversational turns.
See credit plansUS-style hyphenated and toll-free numbers, yes. In our September 2026 checks, 1-800-222-1222 was read as one eight hundred, then the remaining digit groups. International numbers are not reliably grouped. A UK number such as +44 20 7946 0958 can be read as a cardinal.
No. Generation is not a stream. In September 2026, clips under 120 characters averaged 2.65 seconds to a finished MP3 (range 1.46-4.30 seconds). That fits pre-generated prompts, IVR menus, and notification audio, not live conversational turns.
This page covers English voice-agent hard cases: phone numbers, OTP codes, order IDs, email addresses, and a measured latency range. A multilingual text-to-speech page is planned for mixed-language reading. It is not live yet, so this page does not link there.
In live conversation, delays under 150 ms tend to feel immediate, 150-300 ms are usually acceptable, and gaps over 300 ms are noticeable. Those figures describe human perception of a live turn, not a product promise. Miso Voice 2.0 currently returns a finished clip, so use it for prompts you can generate ahead of the call.
US-style hyphenated numbers such as 1-800-222-1222 already grouped correctly in our tests. International numbers are not reliably grouped, so do not assume digit-by-digit reading for those formats until you hear a clip of the exact string.
In our September 2026 checks, mixed codes were spelled character by character by default. XK4P2Q was fully spelled. A7X-93K2-QF was recoverable; the 93 pair was read as ninety-three. You do not need extra markup. Type the code as the caller should hear it.
Everyday sentences are the easy case. Order IDs and confirmation codes are harder because letter-digit grouping is ambiguous. In our September 2026 checks, OTP 482907 and confirmation code XK4P2Q were read character by character. Order ID A7X-93K2-QF was recoverable, with 93 read as ninety-three.
Miso Voice includes built-in text normalization. In our tests, @ became at, dots in an email became dot, and short digit strings such as 482907 were read one by one. International number grouping is not part of that behavior. SSML tags are not supported.
Pick a TTS that can read the strings your agent actually speaks (US-style phone numbers, OTP codes, order IDs, email addresses) and that publishes a measured time to a finished clip. Miso One is the TTS layer for that audio. You still bring telephony, speech-to-text, and dialog. Listen to the clips on this page before you decide.
