
How Unspoken Words Enter AI Meeting Transcripts
A fluent transcript can contain words nobody said. Preserve the recording, inspect silent spans and separate draft text from statements that become meeting evidence.
Read More
Turn Persian voice messages, recordings and meetings into text: phones, Telegram, Bale, Rubika, Word, Google Docs, Iranian services, open models you run, and how to check accuracy.
ZharfAI TeamNovember 16, 202625 minutes
To turn Persian speech into text you have three routes: voice typing on a phone or in a browser when you are the one speaking, a messenger feature or an Iranian service for voice messages and short files, and an open model on your own computer for long recordings and audio that should never leave your hands. According to their official pages, Google Docs voice typing supports Persian, Bale and Rubika convert voice messages to text, and iPhone dictation and Word's Dictate do not support Persian. For Google's and Samsung's keyboards, Telegram and Eitaa no official page settles it, and a thirty-second test on your own phone does.
This guide covers all the ways people in Iran search for this: voice-to-text, voice-message-to-text and speech-to-text. Every product claim here was checked on the product's own official page on November 16, 2026. If you need the opposite direction, reading text aloud, see our Persian text-to-speech guide.
| Where you are | Persian? | Note | Source |
|---|---|---|---|
| Google Docs voice typing | Yes, listed as Farsi | Latest Chrome, Edge or Safari | Docs Help |
| Gboard (Google keyboard) | No official answer | Gboard Help names no languages for standard voice typing; run the thirty-second test | Gboard Help |
| Samsung voice input | No official answer | Samsung's page names no languages; run the test or use Google Voice Typing | Samsung support |
| iPhone dictation | No | Persian is not in the Dictation list | iOS feature availability |
| Dictate in Word | No | Persian is in neither the full nor the preview list | Microsoft support |
| Bale | Yes | Up to 5 minutes per clip and 15 minutes a day | Bale FAQ |
| Rubika | Yes | Latest version; the A icon next to a downloaded voice message | Rubika FAQ |
| Telegram | No official answer | The feature exists, but its official pages name no languages; run the test | Telegram announcement |
| Eitaa | No official answer | The official site describes no such feature; run the test | Eitaa |
| Iranian services | Yes | Online; one also offers an on-premises edition | Iranian services section |
| Open model on your machine | Yes | Offline, on your own hardware | Open models section |
The thirty-second test. Where no official source answers, test it yourself. Set the input language to Persian and say «فردا ساعت ده با آقای رضایی جلسه نداریم» ("we have no meeting with Mr Rezaei at ten tomorrow"). If the text comes out in Persian script with the name, the number and the negation right, it works on your device. If nothing appears, or it comes out in Latin letters, it does not. In a messenger, record that sentence as a voice message, send it to yourself and tap the transcription option next to it.
Gboard Help gives the steps: install Gboard, open an app such as Gmail or Keep, tap where you can enter text, tap the microphone at the top of the keyboard, and speak when you see "Speak now". The page warns that talk-to-text doesn't work with all languages and that some steps work only on Android 7.0 and up.
First add a Persian keyboard; Persian is in Gboard's keyboard language list. On the voice toolbar, which Google describes for tablets, you switch language by tapping the abbreviation for your default language at the bottom left, once several languages are enabled. Spoken punctuation such as "Period" or "Comma" may not be available in every language, per the same page, so add Persian punctuation yourself.
Does it support Persian? Gboard Help gives no language list for standard voice typing. Two newer dictation modes on Pixel phones do list their languages, and Persian is in neither: advanced voice typing covers only English, French, German, Italian, Japanese and Spanish, and Rambler does not name Persian either. The Google Docs language list belongs to Docs, not the keyboard. For standard voice typing, run the thirty-second test on your own phone.
Samsung's support page says Galaxy phones have two built-in voice-to-text tools, Samsung voice input and Google Voice Typing, and that Samsung voice input works only with the Samsung Keyboard. The path it gives for choosing a language:
Samsung notes that screens and settings vary by carrier, software version and model. Does it support Persian? The page names no languages, and we found no other official list of Samsung voice input's offline packs. If Persian appears among the packs on your phone, run the thirty-second test. If it doesn't, the same page gives the alternative: choose Google Voice Typing in that menu and run the test again.
Short answer: iPhone dictation does not support Persian. On Apple's iOS feature availability page the Dictation list goes from Norwegian Bokmål to Polish with no Persian entry. The on-device dictation list and the transcription lists for audio messages in Messages and recordings in Notes have none either. Persian appears on that page only in the keyboard lists, and as "Farsi" under VoiceOver, Live Speech and Read & Speak: the iPhone types Persian and reads Persian text aloud, but does not write down Persian speech.
Practical options on an iPhone are the web version or Telegram bot of Nevisa Live (see the Iranian services section), Google Docs voice typing on a computer, or sending the recording to a computer and transcribing it with an open model. Vira's Avanegar has no iOS version yet, by its own page.
On November 30, 2023 Telegram announced that all users can convert up to 2 messages a week into text by tapping the →A icon next to the message. Telegram Premium removes the cap and, according to the Premium FAQ, covers video messages too. The group-level table on the same page shows that a group reaching boost level 6 unlocks unlimited transcription for its members.
Telegram's API documentation shows that the weekly free count and the maximum message length for free users are set by the server, so the figure of 2 comes from the 2023 announcement and may have changed. Does it support Persian? None of these three official pages names any language. Spend one of your free weekly conversions on the thirty-second test. We did not check buying Premium from Iran and offer no advice on it.
Bale: yes. The Bale FAQ, in its AI section under the heading «تبدیل صوت (وویس) به متن» ("converting voice to text"), says it converts clips up to 5 minutes long, with a daily cap of 15 minutes. The same section says Bale's AI is reached through the @ai_bot bot or the AI tab in the latest Android app. The FAQ does not say exactly where the button is, and it gives no price for the feature.
Rubika: yes. Answering how to turn a voice message into text, the Rubika FAQ says to install the latest version of Rubika, download the voice message and tap the A icon next to it. That answer states no cap or cost.
Like Telegram's, the Bale and Rubika pages name no language. We say "yes" because both are the Persian-language FAQs of Iranian messengers describing the feature to their own users; one run of the thirty-second test puts it beyond doubt.
Eitaa: no official answer. The Eitaa home page, FAQ and download page describe no voice-to-text feature. Send yourself a short voice message and see whether a transcription option appears next to it. If not, save the audio file and transcribe it with one of the file methods further down this page.
Word: no. According to Microsoft's Dictate in Word page, Persian is in neither the fully supported list nor the preview list; Arabic, Hebrew and Turkish are among the preview languages. Dictate needs an active Microsoft 365 subscription and a reliable internet connection, and your speech is sent to Microsoft to produce the text; the page says the service does not store your audio or the transcribed text.
Google Docs: yes. The Docs voice typing help page lists Farsi among the languages that work with voice typing. The same page says your web browser controls the speech-to-text service and sends the text to Docs. The steps, per that page:
Editing and formatting voice commands work only in English and need both the account and the document language set to English; spoken punctuation might not be available in every language. Google's page doesn't say where to choose the voice-typing language, so the first time, run the thirty-second test to make sure Persian comes out. If you don't see the option at all, an administrator may have turned it off for your organisation's accounts. If you work in Word, the practical route is to dictate in Docs and move the text into Word.
We checked four Iranian services on their own sites and did not test any of them with real audio. Every accuracy figure below is the vendor's claim, and none of their pages names a test set.
Read the terms before you upload sensitive audio. Vira's terms give Vira, in one clause, an "absolute" right to use all data it receives from users, at its discretion, to improve its services and develop and upgrade its AI models. Another clause says the data sent for processing and the responses are stored only on the user's hardware and that Vira takes no responsibility for storing data on its servers. A third advises deleting sensitive data from the system archive after processing, and a fourth says user data is not given to third parties except by order of a competent judicial or administrative authority or as the law requires. Read together, these clauses do not make clear where submitted audio is kept or for how long; get a written answer from the vendor before sending customer calls or confidential meetings.
Before using any foreign service, read its terms for a sanctions clause. The one whose terms we read in full is ElevenLabs:
fas) and puts it, for the Scribe v2 model, in its "high accuracy" band of more than 5 and at most 10 percent word error rate. That banding is the vendor's own. Files up to 5 GB and 15 hours are accepted, and Scribe v3 separates up to 32 speakers.So this service is not an option for a user in Iran, and this guide gives no way around that. The legitimate alternatives are an Iranian service above, used after a careful read of its terms, or an open model that runs on your own computer or server so the audio goes nowhere.
An open model has three advantages: the audio never leaves your device, there is no file-length cap and each minute costs nothing. The price is hardware and some command-line work. The table lists the models worth considering for Persian today; the three Persian fine-tunes in it were released in October 2026. Every figure was reported by the model's maker on its own test set, so they are not comparable with each other: test sets and text normalisation differ.
| Model | Reported result | Reported by, on what | Watch out for | Licence |
|---|---|---|---|---|
givia/whisper-large-v3-turbo-persian | 6.4% WER; 11.2% on a 70-minute podcast | Its maker; 300 clips of FLEURS Persian | Also trained on FLEURS-train; outputs punctuation | CC-BY-4.0 |
mehdi-hf/nemotron-asr-streaming-farsi | 8.8% WER; about 26% on conversational speech | Its maker; full FLEURS Persian test (852 clips) | Live streaming; no FLEURS in its training data; no punctuation; numbers as words; Persian only | OpenMDW-1.1 |
MohammadGholizadeh/whisper-large-v3-turbo-persian | 19.4% WER (base model 50.2%) | Its maker; Common Voice 17 Persian test | Trained without timestamps | MIT |
openai/whisper-large-v3, stock | 17.3% WER; turbo 18.7% | The maker of the first row, on the same 300 clips | Base of most Persian fine-tunes | Apache 2.0; turbo under MIT |
| Meta Omnilingual ASR | 3.4% CER | Meta's official table for the 7B model | Character error is not comparable with word error; apart from the Unlimited variants, the models accept only audio shorter than 40 seconds | Apache 2.0 |
Below are two routes you can run yourself today: one simple and multilingual, the other more accurate for Persian.
This works on Windows, macOS and Linux and works through long files in 30-second windows on its own. According to the official Whisper repository you need the ffmpeg tool, and the turbo model needs about 6 GB of GPU memory. Without a GPU it runs on the CPU, just more slowly.
# Install ffmpeg first: sudo apt install ffmpeg on Ubuntu, brew install ffmpeg on macOS
python3 -m venv stt
source stt/bin/activate
pip install -U openai-whisper
whisper voice.ogg --model turbo --language fa --output_format txt
A voice.txt file appears in the folder you ran the command from. Give it a messenger voice message with whatever extension it was saved with, such as ogg; ffmpeg reads it. For subtitles, or to jump back to the moment each sentence was said, use srt instead of txt.
whisper.cpp is a lightweight C++ implementation of Whisper that runs on an ordinary CPU and, on Apple Silicon Macs, uses the GPU through Metal. The maker of the givia model ships a compatible file in that model's Hugging Face repository, about 1.6 GB. Building it needs cmake and a C++ compiler.
# After building, put voice.ogg inside the whisper.cpp folder
git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
cmake -B build
cmake --build build -j --config Release
./models/download-vad-model.sh silero-v6.2.0
curl -L -o ggml-persian-v3-turbo.bin https://huggingface.co/givia/whisper-large-v3-turbo-persian/resolve/main/ggml-persian-v3-turbo.bin
ffmpeg -i voice.ogg -ar 16000 -ac 1 -c:a pcm_s16le voice.wav
./build/bin/whisper-cli -m ggml-persian-v3-turbo.bin -l fa --vad --vad-model ./models/ggml-silero-v6.2.0.bin -otxt -osrt -of voice -f voice.wav
The ffmpeg line is needed because, per the whisper.cpp README, the command-line tool currently reads only 16-bit WAV files. Each option is explained in the whisper-cli README: -l fa sets Persian, -otxt and -osrt write text and subtitle files, and -of voice names them voice.txt and voice.srt. The --vad option with the Silero voice-activity model removes stretches with no speech before transcription. That matters, because the Persian model's maker notes that Whisper models can produce a word such as «موسیقی» ("music") over long stretches of music. If you have an NVIDIA graphics card, the main whisper.cpp README says to install CUDA and build with cmake -B build -DGGML_CUDA=1 as the first build line. The model's maker reports that a 70-minute podcast transcribes in about 25 seconds on one RTX 4070 Ti graphics card; that figure is for a GPU, not a CPU.
If text has to appear while someone is talking, nemotron-asr-streaming-farsi was built for that: 600 million parameters, a look-ahead of 0.32 or 1.12 seconds, and, per its maker, five to ten times faster than real time on an M1 Pro Mac. Its output has no punctuation, writes numbers as words, and drops English words or approximates them in Persian script. Code for live use is in the model card.
A low error rate reported on read sentences will not carry over to your meeting or call. The same model cards show where errors climb:
To measure, take ten minutes of your own real audio: several speakers, some noise, one segment with no speech. Write the correct transcript by hand. Then normalise both the reference and the model output, because «میرود» versus «می رود», or Arabic «ي» versus Persian «ی», should not count as errors. This short script uses the jiwer library to report word and character error rates:
# pip install jiwer
import re
import sys
import jiwer
def normalize(text):
text = text.replace("ي", "ی").replace("ك", "ک") # Arabic yeh and kaf
text = re.sub(r"[ً-ْٰـ]", "", text) # diacritics and tatweel
text = text.translate(str.maketrans("۰۱۲۳۴۵۶۷۸۹٠١٢٣٤٥٦٧٨٩", "01234567890123456789"))
text = text.replace("", " ") # zero-width non-joiner
text = re.sub(r"[^\w\s]", " ", text) # punctuation
return " ".join(text.split())
with open(sys.argv[1], encoding="utf-8") as f:
reference = normalize(f.read())
with open(sys.argv[2], encoding="utf-8") as f:
hypothesis = normalize(f.read())
print("WER:", round(jiwer.wer(reference, hypothesis), 3))
print("CER:", round(jiwer.cer(reference, hypothesis), 3))
Run it as python3 check_wer.py reference.txt voice.txt. The number alone is not enough: check names, numbers, dates and every negation separately against the audio, because one dropped "not" reverses a sentence and disappears inside an average. For the segment with no speech, the correct output is empty; any word there was invented. Fuller metrics for Persian voice products are in our guide to a better Persian voice interface.
Turning a meeting into minutes with AI takes five steps, and only one of them is speech to text:
srt output so each sentence has a start time. Without timestamps a reviewer cannot go back to the audio for that moment.speaker-diarization-community-1 model works out which stretch belongs to which speaker, on your own machine. Per the repository's instructions, ffmpeg must be installed, and to download the model you need a Hugging Face access token and must accept that model's user conditions.The diarisation example, following the pyannote repository:
# pip install pyannote.audio
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-community-1", token="HUGGINGFACE_ACCESS_TOKEN"
)
output = pipeline("meeting.wav")
for turn, speaker in output.speaker_diarization:
print(f"{turn.start:.1f}\t{turn.end:.1f}\t{speaker}")
The speaker labels this prints are not names. Someone who attended has to map them to people, and where two people spoke at once the speaker should stay "unknown". For step four a plain instruction works. Here it is in Persian, followed by its meaning:
این متن پیادهشده یک جلسه است و هر خط زمان شروع و گوینده دارد.
صورتجلسه بنویس با این بخشها: حاضران، تصمیمها، اقدامها (مسئول و مهلت)، موارد باز.
برای هر تصمیم و اقدام زمان همان خط را در پرانتز بیاور.
چیزی را که در متن نیامده ننویس. اگر مسئول یا مهلت گفته نشده، بنویس «گفته نشد».
It says: this is a meeting transcript with a start time and speaker on each line; write minutes with attendees, decisions, actions (owner and deadline) and open items; put the line's time in brackets after every decision and action; do not write anything that is not in the transcript, and write "not stated" when no owner or deadline was given. The timestamps make review short, since each decision can be checked by going to that exact second of the file. For long-term meeting memory, slides and decision follow-up, see our article on the meeting memory.
Call-centre automation usually starts with transcribing recorded calls. Some practical points:
Split the channels. If your phone system records agent and customer on separate channels, transcribe each channel separately and speaker attribution mostly takes care of itself. These two commands turn each channel into a 16 kHz, 16-bit WAV file that whisper-cli can also read. Use one sample call to find out which channel is which:
ffmpeg -i call.wav -af "pan=mono|c0=c0" -ar 16000 -c:a pcm_s16le channel-left.wav
ffmpeg -i call.wav -af "pan=mono|c0=c1" -ar 16000 -c:a pcm_s16le channel-right.wav
Test telephone audio separately. The Whisper and nemotron models on this page work with 16 kHz audio, and the maker of one of the Persian fine-tunes states plainly that noisy telephone audio was not evaluated. Run the accuracy check above on your own calls before deciding anything.
Customer voices are personal data. Callers say names, national ID numbers, card numbers and addresses. Tell callers the call is recorded and transcribed, restrict access to audio and text, set a retention period, and mask sensitive numbers in transcripts. Don't send customer audio to a service whose terms claim the right to use it for training, as Vira's and ElevenLabs' terms do.
Consider your own servers. The open models above run on an organisation's server with no internet connection. If you want models to run on your organisation's own servers so the data stays there, that is what ZharfAI's local AI on your own servers is for.
An AI phone receptionist or AI phone answering service is more than speech to text; it also needs turn-taking, confirmation of amounts and names, and handoff to a person. We cover those in our article on the voice operations layer, and call analysis in our article on voice-of-customer intelligence.
Google Docs voice typing, which officially supports Persian, and an open model on your own computer, which has no cap, are both free. Bale converts up to 15 minutes of voice a day and states no price for it; Rubika describes the feature without stating any cap or cost. The Iranian services offer trial credit or a free mode. Telegram gives a few free conversions a week, but its official pages name no languages.
It depends on your audio. On the FLEURS test, the makers of givia and nemotron report word error rates under 10%; on Common Voice, also read speech, makers report about 19%, and on conversational speech roughly 14% to 31%. Measure ten minutes of your own audio with the script in the accuracy section, then choose.
Yes, in Google Docs voice typing, which officially supports Persian in Chrome, Edge or Safari. For standard Gboard voice typing, Gboard Help gives no language list, and the two newer Pixel dictation modes don't include Persian; confirm with the thirty-second test on your own phone.
The iPhone's own dictation does not support Persian, free or paid. On an iPhone you can use Nevisa Live's web version or Telegram bot, or send the file to a computer and transcribe it with an open model.
Yes. Stock Whisper is multilingual, and Word's Dictate and iPhone dictation support English. The Persian fine-tunes are built for Persian, though: nemotron writes only Persian, and the maker of one Persian Whisper model warns that other languages may have degraded.
Yes, with whisper.cpp or Python Whisper once the model is downloaded. Samsung voice input also has offline language packs, but no official source says Persian is among them.
Rubika can: install the latest version, download the voice message and tap the A icon next to it. Eitaa describes no such feature on its official site; send yourself a short voice message and look for a transcription option, or save the file and transcribe it with one of the file methods on this page.
This guide was reviewed on November 16, 2026 against official pages from Google, Samsung, Apple, Microsoft, Telegram, Bale, Eitaa, Rubika and ElevenLabs, the Iranian services' own sites, and the official model cards and repositories of the open models. Accuracy figures are claims by each model's maker or vendor on their own test sets; we did not re-run any of them. The ffmpeg commands on this page were run on a test file, and the other commands were checked against each repository's official documentation.
TagsPersian Speech to TextVoice to TextPersian ASRSpeech RecognitionWhisperMeeting Minutes

A fluent transcript can contain words nobody said. Preserve the recording, inspect silent spans and separate draft text from statements that become meeting evidence.
Read More
Create a Bale bot in BotFather, run a Python echo bot, connect a language model, move to webhooks or n8n, respect Bale's limits and host the whole bot inside Iran.
Read More
Create a notebook, add sources, check citations, and make Persian audio overviews, quizzes and slides in NotebookLM, now Gemini Notebook, with limits, Iran status and an open-source option.
Read MoreIf 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.