Persian Text to Speech: Methods, Tools and Natural-Sounding Voices

Persian Text to Speech: Methods, Tools and Natural-Sounding Voices

Persian voices on iPhone, Android and Windows, online and Iranian services, open models with runnable code, text preparation for ezafe and numbers, and choosing a voice.

ZharfAI TeamNovember 16, 202636 minutes

To turn Persian text into speech today you have three routes: the Persian voice built into your device (iPhone and Mac have one, Windows does not), an online service that offers Persian voices such as Microsoft Azure, Google's Gemini-TTS, ElevenLabs or the Iranian services Ariana and Avasho, or an open model such as Piper that runs on your own computer without an internet connection. Most free options are for personal listening only: the Mac's system voices, ElevenLabs' free plan and the open Persian voices are not cleared for a video you publish, so for public narration use a paid plan whose terms allow commercial use, or a domestic service once it confirms in writing that you may publish its audio. Whichever you pick, how natural the result sounds depends more on how you prepare the text than on the tool: the unwritten ezafe, numbers and dates, and English words in the middle of a Persian sentence.

This guide was checked on 16 November 2026 (25 Aban 1405) against official pages from Apple, Google, Microsoft, ElevenLabs and OpenAI, the official repositories of the open models, and the Iranian services' own websites. Iranians search for this both as "text to sound" (تبدیل متن به صدا) and "text to speech" (تبدیل متن به گفتار); they mean the same thing. If you need the opposite direction, turning an audio file into text, read our guide to Persian speech to text. We do not rank voices for naturalness here, because the result depends on your text. Instead, we give you a fixed test paragraph so you can judge any tool against the same yardstick.

Which route fits your situation

Your situationSuggested routeCostMain limitation
Listen to text on an iPhone or MacRead & Speak with Apple's Persian voiceFreeAudio files only on the Mac, whose licence limits them to personal, non-commercial use
Listen to text on an Android phonesherpa-onnx speech engine with Piper's Persian voicesFreeAPK installed from the project's site, not from Play
Narrate a video for public releaseA paid plan whose terms allow commercial use, or a domestic service that confirms publishing rightsPaidFree plans, the Mac's voices and the open Persian voices are not cleared for publishing
Phone menu or queue announcementsAzure with SSML, or Ariana on a local serverPaidTest the vendor's claims with your own text
Confidential text or offline workPiper or MMS on your own computerFreePiper's Persian voices are not cleared for commercial use; MMS is non-commercial

Each row is explained, with its sources, in the sections below.

Persian text to speech on iPhone and Mac

Apple's iOS feature availability page lists Persian (as "Farsi") among the languages for VoiceOver, Live Speech and Read & Speak. You do not need a separate app to hear Persian text on an iPhone. Apple does not name its Persian voice on these pages, so take the name from the voice list on your own device.

To turn it on, per Apple's iPhone guide:

  1. Open Settings and go to Accessibility > Read & Speak.
  2. Turn on Speak Selection, then select any text and tap Speak.
  3. Turn on Speak Screen, then swipe down with two fingers from the top of the screen to hear the whole page.
  4. Under Voices, choose a voice and dialect. Default Language sets the language spoken by default, and Detect Languages lets the phone switch language automatically.
  5. Under Pronunciations, record how names and words that are misread should sound.

On older versions such as iOS 18, the guide for that version calls this menu Spoken Content, at Settings > Accessibility > Spoken Content.

Apple's macOS feature availability page lists Farsi for VoiceOver, Live Speech and Read & Speak on the Mac as well. On a Mac, per Apple's guide to having your Mac speak text, go to System Settings > Accessibility > Read & Speak and turn on Speak selection; the default shortcut for reading selected text is Option-Esc. To add a Persian voice, follow Apple's guide to changing the Mac's voice: click the info button next to System voice, pick the language in the sidebar and download the voice. It does not appear in the menu until the download finishes.

The Mac has one advantage over the iPhone: the say command in Terminal writes speech to an audio file. The first command lists installed Persian voices; the second turns the text in matn.txt into a WAV file. Replace VOICE_NAME with a name the first command printed:

bash
say -v '?' | grep ' fa_'
say -v "VOICE_NAME" -f matn.txt -o khoruji.wav --file-format=WAVE --data-format=LEI16@22050

If the first command prints nothing, no Persian voice is installed yet. We ran the second command with an English voice and got a valid 16-bit, 22,050 Hz mono WAV file; no Persian voice was installed on our test machine.

Check the licence before you use that file. Section 2F of Apple's macOS licence allows the system voices only while running macOS and for your own personal, non-commercial projects, and it rules out recording or publishing them in any profit, non-profit, public sharing or commercial context. The file is for your own listening, not for a video you publish.

Free Persian text to speech on Android

Per Google's Android accessibility help, the setting lives at Settings > Accessibility > Text-to-speech output. There you pick the preferred engine, language, speech rate and pitch, and press Play to hear a sample. Install voice data downloads more languages. Google notes that the engine list varies by device and can include Google's engine, the phone maker's engine and third-party engines.

Google's own engine is published on Play as Speech Recognition & Synthesis, and its Play listing says it reads text aloud for Play Books, Google Translate and TalkBack. Neither that listing nor Google's help page publishes the engine's voice languages, so no official page answers whether it has a Persian voice. The thirty-second test settles it on your own phone: in Text-to-speech output, choose Install voice data and look for Persian (فارسی) in the list. If it is there, install it, set the language to Persian and press Play; if it is not, your phone's engine has no Persian voice.

If there is none, the open-source sherpa-onnx project publishes separate installers labelled tts-engine for Persian on its Android speech engine page: Piper voices named fa_IR-amir, fa_IR-gyro, fa_IR-ganji and fa_IR-ganji_adabi, a Mimic3 voice, and a Matcha model that reads Persian and English together. Each comes for four processor architectures (arm64-v8a, armeabi-v7a, x86_64 and x86); most current phones need the arm64-v8a build. After installing, return to Text-to-speech output and choose it as the preferred engine. Google's help only mentions third-party engines from Play, so an engine installed from an APK appearing in that list is expected Android behaviour, not something an official page promises. Download the file only from the project's official page, not from channels or mirror sites. The Persian Piper voices carry licence limits for commercial work, explained in the open-model section.

A third option is Ariana's Android app, described further down; according to its maker, the caller-announcement feature is free and other features draw on account credit.

Windows, Narrator and Edge read aloud

Microsoft's list of Narrator languages and voices for the latest Windows includes Persian neither among the standard voices nor among the natural voices. Windows' built-in screen reader has no Persian voice today.

For Edge's read-aloud, Microsoft's own pages do not settle the question:

  • The Immersive Reader language table marks Persian "No" in both the Web Read Aloud and PDF Read Aloud columns for Chromium-based Microsoft Edge.
  • The Azure Speech language table, covered below, lists two Persian voices, but it describes the Azure service and does not say which voices Edge uses.
  • Microsoft's Read aloud page for Edge publishes no language list, says only a few voices are available offline, and says availability varies by device type, market and browser version.

The thirty-second test: open a Persian web page in Edge, start Read aloud from Settings and more (…) > More tools > Read aloud, the path on that Microsoft page, then open the voice list on the Read aloud bar and look for a Persian voice. If one is listed, play a paragraph; if not, your Edge cannot read Persian. We could not run this test, because Edge is not installed on our review machine.

In Microsoft's web apps the answer is plain: the same Immersive Reader table marks Persian Read Aloud "Yes" in Word for the web, OneNote, Outlook for the web and Teams.

The dependable route on Windows is to run an open model on the same machine. The piper-tts package on PyPI ships a ready-built 64-bit Windows wheel; the steps are below.

Google text to speech online: Translate and Gemini-TTS

Many people paste text into Google Translate and press the speaker button. Google does not publish the list of languages Translate can read aloud on any page we could open. If you see a speaker button next to Persian text, Persian playback works for you.

Google's cloud service is clearer:

  • The older voices have no Persian. The Cloud Text-to-Speech voice list, last updated 10 November 2026, has no fa-IR voice; it covers voice types including Chirp 3: HD, Neural2, WaveNet, Studio and Standard.
  • Persian exists only in Gemini-TTS, as a preview. The Gemini-TTS page lists Persian (Iran), code fa-IR, with launch readiness Preview. It presents 30 voices labelled female or male, for example Kore and Aoede (female) and Charon and Puck (male). Style, pace and emotion are steered with a natural-language prompt; Google's examples are in English.
  • No free tier. The pricing page shows the free usage limit for Gemini-TTS models as "Not available", and billing must be enabled to use the service.
  • Your text is not logged. Google's data logging page says it does not log any customer Cloud TTS text or audio.

We read these Google pages live on 16 November 2026. The pricing page and the Google Cloud terms did not answer direct requests from our machine, and the voice list came back incomplete, so we read those three through a public reader proxy.

Online text to speech services with Persian voices

This table includes only services whose official pages name Persian. The Amazon Polly language list has no Persian, so Polly is not in it.

ServicePersian voices per official pageFree allowanceCommercial useText you submit
Azure AI Speechfa-IR-DilaraNeural (female) and fa-IR-FaridNeural (male)0.5 million characters a month on the free tierLicence terms not reviewed here; Microsoft's responsible-use guidance gives audiobooks and advertising as examples of appropriate useReal-time request text is not stored; batch text is kept in Azure storage, and you can delete it at any time
Google Gemini-TTSfa-IR in preview, 30 voicesNoneNot reviewed separately hereGoogle does not log text or audio
ElevenLabsEleven v3 and Eleven v4 list Persian10,000 credits a monthPaid users onlyLicence to use content to improve the service, with a training opt-out
OpenAIVoices optimised for English; Persian is in the Whisper-based language listNot statedYou must tell listeners the voice is AI-generatedAPI data is not used for training by default; abuse-monitoring logs are kept for up to 30 days by default

Notes on each row:

Azure. The Azure Speech language table has two Persian voices: Dilara (female) and Farid (male). A footnote on both says phonemes, custom lexicons and visemes are not supported for them, which matters in the text-preparation section. The Azure Speech pricing page gives a free allowance of 0.5 million characters a month. Per the text to speech data privacy page, Microsoft does not retain the text of real-time requests or the audio it generates, but scripts you submit for batch long-audio synthesis are stored in Azure storage to process the request, and you can delete them at any time.

ElevenLabs. The ElevenLabs models page lists Persian for Eleven v3, Eleven v4 and their variants, but not for Multilingual v2 or Flash v2.5. The pricing page shows 10,000 credits a month on the free plan, without voice cloning. The ElevenLabs Terms of Service, last updated 9 October 2026, say in section 1(c) that free users may use the services only for non-commercial purposes. Section 4(d) licenses ElevenLabs to use your content to provide and improve the services and to develop new products, and section 4(i) describes opting out of training use from the Data use menu in your account.

OpenAI. The OpenAI text to speech guide says its voices are optimised for English and that language support generally follows Whisper, whose list includes Persian. The same page says OpenAI's usage policies require a clear disclosure to listeners that the voice is AI-generated, and that the tts-1, tts-1-hd and gpt-4o-mini-tts models shut down on 6 January 2027. The retention figures in the table come from the API data page.

Access from Iran to foreign text to speech services

This section states facts with a source for each. It is not legal advice.

  • ElevenLabs names Iran explicitly. Section 15 of the ElevenLabs Terms of Service has users warrant that they are not in a country under U.S. sanctions or embargo, and lists Iran among the examples.
  • OpenAI's supported list omits Iran. OpenAI's supported countries page warns that accessing or offering access from outside the listed countries may lead to account suspension.
  • Azure's account page does not list Iran. The Azure account page says Azure is commercially available in 140 countries and regions, and Iran is not in the list beneath that sentence. We found no Microsoft page in this review that names Iran explicitly.
  • Google Cloud's agreement points to sanctions law. The Google Cloud terms prohibit use that breaches export control laws, which they define to include sanctions administered by the U.S. Treasury's Office of Foreign Assets Control. We saw no Google page naming Iran for this service.
  • Iran is under a U.S. sanctions program. The OFAC Iran sanctions page lists the program's regulations, including the Iranian Transactions and Sanctions Regulations (31 CFR part 560).

We did not test reach or payment from inside Iran and make no claim about either. The practical point is simple: an account opened against a provider's terms can be closed at any moment, taking your files and project with it. For lasting work there are three legitimate routes: the Persian voice on your own device for personal listening, a domestic service that documents its features on its own site, and an open model running on hardware you control.

Afixa and Iranian text to speech services

Afixa. The name Afixa (افیکسا) appears in Google's Persian autocomplete next to "AI text to speech", and the site's own pages do not agree on whether it offers it:

  • Afixa's home page describes voice conversion: you record your own voice with your phone, choose one of more than 100 characters, and your recording is turned into that character's voice, through the Telegram bot @fxa_ir_bot. The About page lists two current services: changing and imitating the voices of speakers and well-known people from a recording the customer makes, and separating the layers of a recording (vocals, piano, drums and so on) from music or film. The site footer names the company as Neda Artificial Intelligence.
  • Several article pages on the same site describe text to speech. One calls Afixa a free Persian text-to-speech bot with a choice of speaker voice, while another says the text-to-speech service "will soon be added to Afixa on a trial basis".

The articles carry no dates, and we did not test the bot, so we cannot say which description is current; ask the company before you rely on it for text to speech. The About page also says the company is in contact with artists, dubbing actors and voice actors to model their voices. That is a vendor claim, and if you intend to publish a specific person's voice, you need that person's written consent (see the section on cloning).

Avasho. The Part AI Research Center describes Avasho on its product page as an API that reads Persian text aloud for developers to build into their software. Its claims, which we have not tested: a choice of male or female voice, diacritics you can add to steer the reading, adjustable speed, more than 300,000 words in its pronunciation vocabulary, and a new voice built from ten minutes of a chosen speaker's recordings. The page gives no prices or free allowance, and the separate Avasho site it links to did not answer our requests. Building a voice from someone's recordings is voice cloning, so the consent rules below apply.

Ariana. Asr Gooyesh Pardaz presents its Ariana text to speech service on the FarsiReader site. The following are the vendor's own claims, which we have not tested:

  • it reads diacritics, numbers, symbols and English words, and lets users add words;
  • it outputs mp3, wave and ogg:opus and connects over REST and WebSocket;
  • it can be installed offline on your own server or used through prepaid credit on the company's servers, with free starting credit after sign-up;
  • per the installation page, the offline edition runs on Windows Vista or later and Windows Server 2008 or later and needs 512 MB of free memory and a USB hardware key; the home page FAQ adds that a "licence" means one concurrent request, not one installation;
  • per the products page, there is a Telegram bot, an Android app, a WordPress plugin and a cloud client; the home page also lists a Bale bot;
  • per the applications page, it connects to the Asterisk, Elastix and FreeSWITCH telephony platforms and can read in a custom voice recorded for the customer.

The FarsiReader site's copyright line shows 2022; before buying, ask the vendor for the current version and for a sample made from the test paragraph below.

Open models: Persian text to speech on your own computer

An open model means the text never leaves your computer and you need neither an account nor an internet connection. But open source does not mean free for commercial use; take the licence column seriously.

ModelPersian voicesLicenceHardware
PiperSix fa_IR voicesEngine GPL-3.0; no Persian voice is on Piper's commercial-use listOrdinary CPU; each voice about 63 MB
Meta MMS-TTSOne voice (facebook/mms-tts-fas)CC-BY-NC 4.0, non-commercial onlyOrdinary CPU; model file about 145 MB
OmniVoicePersian among more than 600 languagesCode Apache-2.0, model weights CC-BY-NCGPU, or Apple silicon

Piper: fast, light and offline

Piper is a local text to speech engine that uses espeak-ng to turn text into phonemes, released under GPL-3.0. Version 1.8.0 of its package on PyPI has ready-built wheels for 64-bit Windows, macOS (Intel and Apple silicon) and Linux, and needs Python 3.9 or later. The Piper voices repository has six Persian voices: amir, ganji, ganji_adabi, gyro, magnoliasis and reza_ibrahim, all at medium quality and a 22,050 Hz sample rate.

Following Piper's command-line guide, these commands produce a WAV file on macOS and Linux:

bash
python3 -m venv tts-env
source tts-env/bin/activate
pip install piper-tts
python3 -m piper.download_voices fa_IR-amir-medium
python3 -m piper -m fa_IR-amir-medium -f salam.wav -- 'سلام. این یک آزمایش صدای فارسی است.'

The fourth command downloads the voice into the current folder and the last one writes salam.wav. Running python3 -m piper.download_voices without a voice name lists every voice. The --input-file option reads the text from a file, --sentence-silence 0.5 adds half a second of silence after every sentence except the last, and --length-scale 1.15 slows the speech slightly. No GPU is needed; if you have an NVIDIA card, --cuda with the onnxruntime-gpu package puts it to use.

On Windows, replace the second line with tts-env\Scripts\activate, and use python if python3 is not recognised. The Windows command prompt does not treat single quotes as quotes, so save the Persian text in a UTF-8 text file and pass it with --input-file matn.txt instead of putting it on the command line. We did not test this on Windows.

The same in Python, with a speed setting, following Piper's Python API guide:

Python
import wave

from piper import PiperVoice, SynthesisConfig

voice = PiperVoice.load("fa_IR-amir-medium.onnx")
config = SynthesisConfig(length_scale=1.15)  # slightly slower than the default

with wave.open("salam.wav", "wb") as wav_file:
    voice.synthesize_wav("سلام. این یک آزمایش صدای فارسی است.", wav_file, syn_config=config)

A length_scale above 1 slows speech down; a slightly slower pace usually helps with instructional narration.

Licensing of Piper's Persian voices. The project's own list of voices cleared for commercial use says Piper grants no licence over voice models and that a voice's rights come from its training data and whoever trained it. No Persian voice is on the list: ganji and ganji_adabi derive from amir, which was itself fine-tuned from the English lessac voice, and the magnoliasis model card states that its dataset has no licence at all and should not be assumed usable for any purpose. The amir and reza_ibrahim cards describe their own recordings as CC0, but both were also fine-tuned from lessac, and the gyro card only points to a URL for its licence, a group the Piper list calls not yet reviewed. The same list says lessac-derived voices are being retrained over time and that a copy downloaded before retraining does not qualify, so reread the list and your copy's model card before any commercial use. Since the magnoliasis card grants no rights for any use, leave that voice out; before publishing with the others, take legal advice or build a dedicated voice with a speaker's written consent.

MMS-TTS: one Persian voice from Meta

The facebook/mms-tts-fas model card describes it as part of Meta's Massively Multilingual Speech project, a VITS model under CC-BY-NC 4.0. It outputs 16 kHz audio. Following the Transformers VITS documentation, the code below generates one sentence and writes it to a WAV file:

bash
pip install transformers torch scipy
Python
import scipy.io.wavfile
import torch
from transformers import AutoTokenizer, VitsModel, set_seed

model = VitsModel.from_pretrained("facebook/mms-tts-fas")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-fas")

text = "امروز بیست و پنجم آبان است و جلسه ساعت چهارده و سی دقیقه شروع می‌شود"
inputs = tokenizer(text, return_tensors="pt")

set_seed(555)  # the output is random; a fixed seed makes it repeatable
with torch.no_grad():
    waveform = model(**inputs).waveform[0]

scipy.io.wavfile.write("mms-fa.wav", rate=model.config.sampling_rate, data=waveform.cpu().numpy())

One practical point the model card leaves out: this model's vocab.json has only 44 symbols, made up of Persian and Arabic letters, the space, the hyphen, the underscore, the apostrophe and the hamza that sits above ه. Digits, Latin letters, the half-space and the short-vowel marks (zabar, zir and pish) are not in it. The Transformers documentation also says these tokenizers remove punctuation by default. So for this model, write numbers as words, write English words in Persian letters, and generate sentences separately to get pauses. Adding diacritics will not help it either.

OmniVoice: many languages, but a GPU and a non-commercial licence

The OmniVoice repository, from the same k2-fsa group behind sherpa-onnx, offers a model covering more than 600 languages; its language list includes Persian, code fa, with 366 hours of training data. The OmniVoice model card puts the code under Apache 2.0 and the trained model under CC-BY-NC because of restrictions in its training data. It runs on PyTorch with a GPU or Apple silicon; the repository states no minimum memory, and its published tests ran on an H100 and on Intel Arc cards with 4 GB and 16 GB. It can "design" a voice from attributes such as gender, age and pitch without cloning anyone, but the repository notes that voice design was trained on Chinese and English data only and may be unstable in other languages. Its disclaimer prohibits unauthorised voice cloning, voice impersonation and fraud. Given the licence and the hardware, Piper is the practical open route for most readers, so we give no OmniVoice setup here.

The code in this section has been syntax-checked but not executed; installing the packages was not possible on our review machine.

Why Persian is hard for a machine voice

Before a synthetic voice can make sound it has to work out how each word is pronounced, and Persian script leaves much of that unwritten. Six problems, each with a sentence you can test yourself:

  1. The ezafe is not written. "رئیس اداره کل امور مالی استان" (head of the provincial general finance department) carries five ezafe vowels and none is visible. If the voice drops them, the phrase sounds like a list of separate words.
  2. Words spelled alike. "کشتی" is both kashti (ship) and koshti (wrestling); "مرد" is both mard (man) and mord (died). Try: "پدربزرگ در جوانی کشتی می‌گرفت و در پیری با کشتی به جنوب رفت." (Grandfather wrestled when young and in old age went south by ship.)
  3. Short vowels are not written. "ملک" can be malek, melk, malak or molk, and only the sentence decides.
  4. Numbers, dates and units. "۱۴۰۵/۰۸/۲۵", "ساعت ۱۴:۳۰", "۱۲٫۵٪", "قرن ۲۱" and "۳۷۰ میلیون تومان" each have several possible readings: ordinal or cardinal, digit by digit or as a whole, with or without the unit.
  5. Persian and English in one sentence. "فایل PDF راهنما را روی iPhone باز کنید" switches writing direction and hides the ezafe after PDF. Some models skip Latin words altogether.
  6. Punctuation and the half-space. Pauses and question intonation come from commas and question marks. "می‌خواهم" with a zero-width non-joiner, "میخواهم" joined and "می خواهم" with a full space are three different strings to a machine, and Arabic "ي" and "ك" are not the same characters as Persian "ی" and "ک".

We do not claim that any particular tool fails on these; they are the points you should test. Our article on better Persian voice interfaces explains why Persian text normalisation is a product decision and why speech synthesis must be evaluated separately from speech recognition.

Preparing text before converting it to speech

Most of the quality gain comes from this step, not from switching tools.

  1. Normalise characters. Replace Arabic forms with Persian "ی" and "ک" and put half-spaces where they belong.
  2. Write numbers the way they should be read. Write "۲۵ آبان ۱۴۰۵" as "بیست‌وپنجم آبان هزار و چهارصد و پنج" and "۱۴:۳۰" as "ساعت چهارده و سی دقیقه" or "دو و نیم بعدازظهر". Separate the digits of phone numbers and codes with spaces or commas.
  3. Add diacritics only where there is ambiguity. "کُشتی" versus "کَشتی", "مُلک", and the ezafe kasra in "گزارشِ فروش". Then listen to whether your tool uses them; MMS's vocabulary, for example, has no diacritics at all.
  4. Decide how English words are read. If the tool reads only Persian, write "PDF" as "پی‌دی‌اف" and "iPhone" as "آیفون".
  5. Split long sentences. One idea per sentence. A sentence that leaves you breathless when read aloud also leaves a synthetic voice flat.
  6. Make pauses with punctuation. A comma for a short pause, a full stop for a full one, and a question mark for rising intonation.
  7. Keep a pronunciation list. Record names, brands and technical terms once with the correct pronunciation: Pronunciations on iPhone and Mac, Ariana's add-a-word feature (per the vendor), or a replacement table in your own code before the text is sent.

SSML in Azure: what works for Persian

SSML is a standard markup that tells a speech engine where to pause, how fast to speak and what to say. In Azure, Persian comes with limits. Per the language table, the phoneme element and custom lexicons are not supported for the Persian voices, and per the SSML pronunciation documentation, only the characters and spell-out values of say-as work in every language; the others, such as date and cardinal, are limited to a list of languages that does not include Persian. What remains: break for pauses of up to 20 seconds per the SSML structure documentation, prosody for rate and pitch, which per the SSML voice documentation changes speed between half and double, and sub to speak substitute text:

XML
<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="fa-IR">
  <voice name="fa-IR-DilaraNeural">
    <prosody rate="-10%">
      گزارشِ فروشِ این فصل
      <break time="400ms"/>
      در قالب <sub alias="پی‌دی‌اف">PDF</sub> برای همه فرستاده شد.
    </prosody>
  </voice>
</speak>

So in Azure, too, you write Persian numbers and dates out in words yourself. We did not review SSML support for Gemini-TTS or ElevenLabs and make no claim about it.

Fixing pronunciation in Piper

Piper's command-line guide lets you put espeak-ng phonemes directly into the text inside [[ ]]. To see the phonemes espeak-ng produces for a Persian phrase, if you have installed espeak-ng separately:

bash
espeak-ng -v fa --ipa=3 -q "گزارشِ فروش"

The espeak-ng language list has Persian under the identifier fa, and the configuration file of the fa_IR-amir-medium voice uses the same identifier. Edit the output and place it inside [[ ]] instead of the problem word.

Choosing a female voice or a narrator's voice

A female voice. In Azure, fa-IR-DilaraNeural is female and fa-IR-FaridNeural male. Google's Gemini-TTS page labels the gender of each of its 30 voices. The model cards for Piper's Persian voices mostly do not state gender, except gyro, whose dataset is described as male; listen to samples before choosing. OmniVoice can design a female voice, with the caveat above about voice design outside Chinese and English. According to their makers, Ariana's Telegram bot lets you choose the voice type and gender, and Avasho offers a male or a female voice.

A narrator's voice. What people call a "narrator's voice" is usually a slightly slower pace, steady pitch and well-placed pauses. Get close to it with prosody in Azure, length_scale in Piper or a prompt in Gemini-TTS, and build pauses with punctuation.

What does "natural" mean? A natural voice pronounces words correctly, voices the ezafe, distinguishes statement and question intonation, stays consistent from one sentence to the next and has no clicks or cut-offs. None of that can be judged from a single short sentence.

A fixed test paragraph

Give this paragraph, unchanged, to every tool you evaluate:

روز ۲۵ آبان ۱۴۰۵، ساعت ۱۴:۳۰، خانم کرمی گزارش فروش سه‌ماهه را خواند: «درآمد ۱۲٫۵ درصد بیشتر شد و ۳۷۰ میلیون تومان به حساب شرکت نشست.» بعد گفت مرد جوانی که دیروز زنگ زد، می‌خواهد فایل PDF راهنما را روی iPhone و Android امتحان کند. پدربزرگش در جوانی کشتی می‌گرفت و سال‌ها بعد با کشتی به بندرعباس رفت. آیا جلسه بعدی سه‌شنبه است؟ نه؛ پنج‌شنبه!

In English: on 25 Aban 1405 at 14:30, Ms Karami read the quarterly sales report: revenue rose 12.5 percent and 370 million tomans reached the company account. She added that a young man who called yesterday wants to try the PDF guide on iPhone and Android. His grandfather wrestled when young and years later sailed by ship to Bandar Abbas. Is the next meeting on Tuesday? No; Thursday!

What you should hearWhere it is in the paragraph
Ordinal date and full year"بیست‌وپنجم آبان هزار و چهارصد و پنج"
Time"چهارده و سی دقیقه", not "one, four, thirty"
Decimal and amount"دوازده و نیم درصد" or "دوازده ممیز پنج درصد", "سیصد و هفتاد میلیون تومان"
Ezafe"گزارشِ فروشِ سه‌ماهه", "مردِ جوان", "فایلِ PDFِ راهنما"
Words spelled alike"کُشتی می‌گرفت" (wrestled) and "با کَشتی" (by ship)
English wordsPDF, iPhone and Android read aloud, not dropped
Rhythm and pausesPauses after the colon and semicolon, rising question, exclamation

For a fair comparison, listen to each output once on a phone speaker and once on headphones, hide the tool's name from listeners, and ask three to five Persian speakers to score each output from 1 to 5 and count pronunciation errors. Then rewrite the paragraph using the preparation steps above and test again; the difference between the two versions shows how much of the problem was in the text and how much in the tool.

Uses: video narration, audiobooks, phone menus and accessibility

Video narration. Settle the licence before you generate anything you will publish; the FAQ below sums up what each option's licence says. Generate scene by scene so that fixing one sentence does not mean regenerating the whole file, and keep the WAV output so the editor loses no quality. If listeners might take the voice for a real person, disclose that it is synthetic; OpenAI makes this mandatory.

Audiobooks and articles. Generate long text chapter by chapter. For audio longer than 10 minutes, such as audiobooks and lectures, Microsoft points to its batch synthesis route, which, unlike real-time synthesis, stores your text in Azure storage. The rights to the text itself are yours to secure: Microsoft states that obtaining the licences for content you send to the service is the customer's responsibility.

Phone menus and call centres. In a phone menu, numbers, amounts and reference codes are the parts that matter most; write them out in words with pauses, and listen on a real phone line, not just headphones. Ariana says it connects to Asterisk and FreeSWITCH. If the system is meant to hold a conversation, the voice is only one layer; our article on voice agents in production covers turn-taking, confirmation and handoff to a human.

Accessibility for blind and low-vision readers. VoiceOver reads Persian on iPhone and Mac; Windows Narrator has no Persian voice. On Android, Google's Play listing names TalkBack among the apps that speak through the system engine, so after installing a Persian engine, check that TalkBack reads Persian with it. If you are making a site or app speak for blind users, let the listener control the speed, because every screen-reader user has their own preferred rate.

Recent models such as OmniVoice imitate a person's voice from a reference recording without any separate training, which also makes voice cloning a tool for fraud: a call in a manager's voice ordering a transfer, or a message in a relative's voice asking for money. Here is what the tools' own terms say:

  • ElevenLabs: section 3 of its Prohibited Use Policy bans cloning, synthesising or impersonating the voice, identity or likeness of any living person without their explicit and informed consent or other valid authorisation, and explicitly bans fraud. Section 10 requires you to disclose that people are interacting with AI where the law requires it or where not doing so would materially mislead them, and section 8 prohibits circumventing voice verification systems.
  • Microsoft: per its disclosure for voice talent, building a custom voice requires the voice talent's explicit written permission. The talent must record an acknowledgement statement in their own voice, which Microsoft may compare with the training audio using speaker verification. The same page lists impersonation without explicit permission and unsolicited calls among inappropriate uses. In any case, the Azure language table lists Persian for neither professional nor personal voice.
  • OmniVoice: the repository's disclaimer prohibits unauthorised cloning, voice impersonation and fraud.

Our working rule: do not clone anyone's voice without written consent. The consent should state the use, duration, media and the right to withdraw. Label every synthetic recording you publish, and keep the consent form and the generated scripts. Agree a spoken code word in your family and your organisation for confirming money requests made by phone. How to check whether an audio or video file is authentic is covered in our article on synthetic media and content authenticity, and what it means for singers and artists in AI in the music industry.

Organisations and confidential text: generating speech on your own servers

Before sending a contract, a patient file, a financial report or an internal memo to any online service, ask one question: where does this text go, and how long does it stay there? The "Text you submit" column of the services table above gives each provider's own answer.

If none of these fits your data policy, generate speech on your organisation's own servers:

  1. An open model on an internal server. Piper runs on an ordinary CPU and the text never leaves your network. Check the licence limits of the Persian voices with legal counsel.
  2. Domestic software installed offline. According to its maker, Ariana runs on the organisation's Windows server without an internet connection.
  3. Your organisation's own voice. Contract a professional speaker, obtain written consent and train a dedicated voice. For a sense of the work involved, Microsoft asks voice talent for at least 300 recorded lines for a proof of concept and about 2,000 for a production voice on its own service.

Before rollout, extend the test paragraph with the organisation's real text, build a pronunciation list for its terms, set how long audio files are kept, and make someone responsible for listening to random samples every week. If your organisation also wants to run language models inside its own network, that is what our private AI on your own servers service covers.

Frequently asked questions

Is there free AI text to speech for Persian?

Yes, for personal use. The Persian voice on iPhone and Mac is free, Piper and MMS run free on your own computer, Azure gives 0.5 million free characters a month and ElevenLabs 10,000 credits a month. Google's Gemini-TTS has no free tier. The Mac's system voices, ElevenLabs' free plan, MMS and OmniVoice are limited to non-commercial use.

Does Google text to speech read Persian?

In Google's cloud service, Persian is listed only for Gemini-TTS, as a preview, and the older voices have no Persian. For Google Translate and the Android speech engine, Google has not published an official language list on any page we opened; test on your own device.

How do I get free Persian text to speech on Android?

Look for Persian under Settings > Accessibility > Text-to-speech output and Install voice data. If it is not there, install one of sherpa-onnx's Persian engines from the project's official page and set it as the preferred engine.

Which is the best online Persian text to speech site?

There is no fixed answer, because quality depends on your text. Give this article's test paragraph to two or three candidates, score the results with a few listeners, and read each one's commercial licence and text-retention policy before choosing.

Does Afixa convert text to speech?

Its own pages disagree. The home and About pages describe only voice conversion from your own recording and separating audio layers; some articles on the site describe a free text-to-speech bot, and another says text to speech will soon be added on a trial basis. We did not test the bot, so ask the company before relying on it.

Can I publish the generated voice on YouTube or in an advert?

That depends on the tool's licence. Apple's macOS licence rules out publishing its system voices, ElevenLabs allows commercial use for paid users only, MMS and OmniVoice are non-commercial, and Piper's Persian voices are not on the project's commercial-use list. Microsoft's responsible-use guidance gives audiobooks and advertising as examples of appropriate use, but we did not review Azure's or Google's licence terms here, and the Iranian vendors' pages say nothing about rights to the audio, so ask before you publish. In every case, do not publish a real person's voice without their written consent.

Why does the Persian voice skip the ezafe?

Because the ezafe is not written in Persian script and the tool has to guess it. Write the kasra where it matters, listen to whether the tool uses it, and if not, rewrite the sentence so the ezafe chain is shorter.

Source Notes

This guide was reviewed on 16 November 2026 (25 Aban 1405) against official pages from Apple, Google, Microsoft, Amazon, ElevenLabs and OpenAI, the OFAC Iran sanctions page, the official repositories and model cards for Piper, MMS-TTS, OmniVoice, sherpa-onnx and espeak-ng, and the Afixa, Ariana and Part AI websites. Every page was read live on that date; Google's Text-to-Speech pricing page, the Google Cloud terms and Google's Android help page did not answer direct requests from our machine, and the Cloud TTS voice list came back incomplete, so those four were read through a public reader proxy. Prices, allowances and voice lists are vendor statements and change quickly. For the Iranian services we quote only the claims on their own sites. We did not rank any voice for quality. The say command was run with an English voice; the Python code was syntax-checked but not executed.

TagsPersian text to speechFarsi TTSspeech synthesisPersian voicePiperSSMLaccessibility

Related Posts

Run this inside your own team

If you want the agents and automation in this guide working for your engineering team or business processes, start with a small pilot you can measure.