Persian Speech Recognition

Automatic speech recognition, commonly abbreviated ASR, converts speech into written text. It should be distinguished from speaker recognition, which concerns who is speaking. A resource can support both tasks, but success at identifying a voice does not establish that its words have been transcribed correctly. DeepMine, a Persian and English speech database, explicitly brings these research purposes together while distinguishing their experimental settings. s1

For Persian, the usefulness of a speech resource depends on more than its language label. DeepMine's authors discuss metadata covering attributes such as province, age, and education, alongside accent variation. These descriptions help researchers investigate differences among speakers; they do not establish that a trained system performs equally well for every community. Persian-language findings should therefore be read with attention to the people and recording conditions represented in the experiment. s1

[s1]

Persian continuous-speech research predates the present generation of large neural models. A paper presented at Eurospeech in 1999 described adapting a phonetically balanced Persian database for recognition experiments. Its author found that results changed dramatically when the recognizer could rely on a restrictive word-pair grammar. The paper attributed the unusually strong constrained result to the predictability of the database rather than treating it as unrestricted conversational competence. s2

That historical example remains useful because it exposes a general evaluation problem: an apparently impressive result can partly reflect an easier test. A benchmark's vocabulary, sentence construction, and permitted context shape what its score means. The paper's conclusions concern its own database and recognizer, but the distinction between a constrained experiment and open-ended speech is still essential when interpreting later claims. s2

[s2]

FLEURS provides another route into Persian speech research. It is a multilingual speech benchmark built from parallel text, and its dataset documentation includes an Iranian Persian configuration, fa_ir. Separate training, validation, and test partitions make it possible to evaluate systems under a shared structure rather than choosing a new collection for every experiment. s3

The documentation also distinguishes training speakers from evaluation speakers. This matters because recognizing new recordings from familiar voices is a different challenge from handling previously unheard speakers. Nevertheless, a common benchmark is a defined sample, not a substitute for testing the conditions of a particular application. Its Persian configuration should not be interpreted as comprehensive coverage of every Persian variety or every language spoken in Iran. s3

[s3]

The Neyshekar preprint, submitted in September 2026, describes an open Persian read-speech corpus intended to cover formal and informal language, named entities, and longer utterances. Its abstract reports evaluation with different ASR architectures and an independent public sample. Importantly, it also states that transfer and mixture benefits were not consistent across architectures and training budgets. The reported improvements therefore belong to specified experiments, rather than proving that adding the corpus always improves recognition. s4

For readers assessing the field, these resources point toward a practical conclusion: progress depends on better matched evidence as well as larger models. Historical experiments reveal the effects of restrictive tests; multilingual benchmarks support shared comparisons; newer Persian corpora explore coverage gaps. None of these sources alone measures national usage or establishes the reliability of a deployed transcription service. s2s3s4

[s2] [s3] [s4]