Dataset Viewer
Auto-converted to Parquet Duplicate
language
stringclasses
7 values
voice_id
stringlengths
8
30
script_version
stringclasses
3 values
result
stringclasses
2 values
audio_seconds
float64
3.7
16.5
βŒ€
native_words_recognised_pct
float64
0
100
βŒ€
english_loanwords_recognised_pct
float64
0
100
βŒ€
urdu
ur-PK-UzmaNeural
native
ok
12.9
100
100
urdu
ur-PK-UzmaNeural
mixed
no audio
null
null
null
urdu
ur-PK-UzmaNeural
roman
no audio
null
null
null
urdu
ur-PK-AsadNeural
native
no audio
null
null
null
urdu
ur-PK-AsadNeural
mixed
ok
13.1
100
100
urdu
ur-PK-AsadNeural
roman
no audio
null
null
null
urdu
ur-IN-GulNeural
native
ok
14.4
88
100
urdu
ur-IN-GulNeural
mixed
ok
14.2
92
100
urdu
ur-IN-GulNeural
roman
ok
10.8
28
100
urdu
ur-IN-SalmanNeural
native
ok
14.3
100
100
urdu
ur-IN-SalmanNeural
mixed
ok
13.9
100
100
urdu
ur-IN-SalmanNeural
roman
ok
10.7
20
100
urdu
en-IN-NeerjaNeural
native
no audio
null
null
null
urdu
en-IN-NeerjaNeural
mixed
ok
6.1
0
89
urdu
en-IN-NeerjaNeural
roman
ok
15.3
100
100
urdu
en-US-AriaNeural
native
no audio
null
null
null
urdu
en-US-AriaNeural
mixed
ok
6
0
100
urdu
en-US-AriaNeural
roman
ok
15.8
12
0
urdu
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
urdu
en-US-AndrewMultilingualNeural
mixed
no audio
null
null
null
urdu
en-US-AndrewMultilingualNeural
roman
ok
14.1
52
89
bengali
bn-BD-NabanitaNeural
native
ok
12.9
88
78
bengali
bn-BD-NabanitaNeural
mixed
no audio
null
null
null
bengali
bn-BD-NabanitaNeural
roman
ok
13
0
0
bengali
bn-BD-PradeepNeural
native
ok
12.9
88
56
bengali
bn-BD-PradeepNeural
mixed
ok
12.6
76
44
bengali
bn-BD-PradeepNeural
roman
ok
12.6
0
11
bengali
bn-IN-TanishaaNeural
native
ok
14.4
41
22
bengali
bn-IN-TanishaaNeural
mixed
ok
14.2
76
33
bengali
bn-IN-TanishaaNeural
roman
ok
14.1
0
0
bengali
bn-IN-BashkarNeural
native
ok
13.4
100
89
bengali
bn-IN-BashkarNeural
mixed
ok
13.2
53
11
bengali
bn-IN-BashkarNeural
roman
ok
12
0
0
bengali
en-IN-NeerjaNeural
native
no audio
null
null
null
bengali
en-IN-NeerjaNeural
mixed
ok
6.1
0
0
bengali
en-IN-NeerjaNeural
roman
ok
15.7
35
11
bengali
en-US-AriaNeural
native
no audio
null
null
null
bengali
en-US-AriaNeural
mixed
ok
6
0
100
bengali
en-US-AriaNeural
roman
ok
15.5
0
0
bengali
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
bengali
en-US-AndrewMultilingualNeural
mixed
no audio
null
null
null
bengali
en-US-AndrewMultilingualNeural
roman
no audio
null
null
null
persian
fa-IR-DilaraNeural
native
ok
12.8
100
100
persian
fa-IR-DilaraNeural
mixed
ok
12.4
100
17
persian
fa-IR-DilaraNeural
roman
ok
16
0
50
persian
fa-IR-FaridNeural
native
ok
13.1
100
100
persian
fa-IR-FaridNeural
mixed
ok
12.2
100
83
persian
fa-IR-FaridNeural
roman
ok
12.9
10
0
persian
en-IN-NeerjaNeural
native
no audio
null
null
null
persian
en-IN-NeerjaNeural
mixed
ok
5
10
50
persian
en-IN-NeerjaNeural
roman
ok
14.6
67
100
persian
en-US-AriaNeural
native
no audio
null
null
null
persian
en-US-AriaNeural
mixed
ok
5.1
10
33
persian
en-US-AriaNeural
roman
ok
14.3
76
83
persian
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
persian
en-US-AndrewMultilingualNeural
mixed
ok
11.2
95
17
persian
en-US-AndrewMultilingualNeural
roman
no audio
null
null
null
arabic
ar-SA-ZariyahNeural
native
ok
14.5
100
100
arabic
ar-SA-ZariyahNeural
mixed
ok
14.6
100
100
arabic
ar-SA-HamedNeural
native
ok
14.5
100
100
arabic
ar-SA-HamedNeural
mixed
ok
14.3
100
100
arabic
ar-EG-SalmaNeural
native
ok
14.8
100
100
arabic
ar-EG-SalmaNeural
mixed
ok
14.7
100
100
arabic
en-IN-NeerjaNeural
native
no audio
null
null
null
arabic
en-IN-NeerjaNeural
mixed
ok
3.7
0
100
arabic
en-US-AriaNeural
native
no audio
null
null
null
arabic
en-US-AriaNeural
mixed
ok
3.9
0
100
arabic
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
arabic
en-US-AndrewMultilingualNeural
mixed
no audio
null
null
null
tamil
ta-IN-PallaviNeural
native
ok
15
92
67
tamil
ta-IN-PallaviNeural
mixed
ok
12.9
58
33
tamil
ta-IN-ValluvarNeural
native
ok
14.9
83
56
tamil
ta-IN-ValluvarNeural
mixed
ok
12.8
92
56
tamil
ta-LK-KumarNeural
native
ok
15.7
58
67
tamil
ta-LK-KumarNeural
mixed
ok
13.5
92
44
tamil
en-IN-NeerjaNeural
native
no audio
null
null
null
tamil
en-IN-NeerjaNeural
mixed
ok
6.3
0
0
tamil
en-US-AriaNeural
native
no audio
null
null
null
tamil
en-US-AriaNeural
mixed
ok
6.2
0
0
tamil
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
tamil
en-US-AndrewMultilingualNeural
mixed
no audio
null
null
null
telugu
te-IN-ShrutiNeural
native
ok
13.9
85
89
telugu
te-IN-ShrutiNeural
mixed
ok
13.4
54
33
telugu
te-IN-MohanNeural
native
ok
14.2
62
78
telugu
te-IN-MohanNeural
mixed
ok
13.6
85
56
telugu
en-IN-NeerjaNeural
native
no audio
null
null
null
telugu
en-IN-NeerjaNeural
mixed
ok
6.3
0
89
telugu
en-US-AriaNeural
native
no audio
null
null
null
telugu
en-US-AriaNeural
mixed
ok
6.2
0
89
telugu
en-US-AndrewMultilingualNeural
native
no audio
null
null
null
telugu
en-US-AndrewMultilingualNeural
mixed
no audio
null
null
null
hindi
hi-IN-SwaraNeural
mixed
ok
15.8
82
44
hindi
hi-IN-SwaraNeural
native
ok
16.2
88
56
hindi
hi-IN-SwaraNeural
roman
ok
15.6
82
44
hindi
hi-IN-MadhurNeural
mixed
ok
15.5
88
56
hindi
hi-IN-MadhurNeural
native
ok
15.2
97
89
hindi
hi-IN-MadhurNeural
roman
ok
15.5
91
67
hindi
en-IN-NeerjaNeural
mixed
ok
6.1
9
33
hindi
en-IN-NeerjaNeural
native
no audio
null
null
null
hindi
en-IN-NeerjaNeural
roman
ok
14.8
82
44
End of preview. Expand in Data Studio

SpeakScribe open measurements

Measurements behind the voice and transcription pages on speakscribe.me (also on GitHub), published so anyone can check, reuse or compare them. Licence: CC BY 4.0 β€” use freely, credit "SpeakScribe (speakscribe.me)".

File What it is Rows
voices_measured_2026_10.csv Every text-to-speech voice on SpeakScribe (Studio = Kokoro, Global = Microsoft neural): median pitch, intonation range, spectral centroid ("brightness") and reading pace 361 voices, 75 languages
transcription_accuracy_fleurs_2026_10.csv Whisper large-v3-turbo transcription accuracy per language on Google's FLEURS test set (20 clips per language) 40 languages
mixed_script_tts_2026_11.csv Can a voice read a script that mixes its language with English words, or that is typed in English letters (Roman Urdu, Banglish, Finglish, Hinglish)? 115 voice Γ— script readings, 7 languages

How it was measured

Voices (22–30 October 2026). Each voice's sample MP3 (the same sentence for every voice of a language and engine), decoded to 16 kHz mono. Pitch: median of 40 ms autocorrelation frames (60–400 Hz); for English voices, the median pitch measured with Praat on a longer 59-word passage. Intonation range: 10th–90th percentile spread, in semitones. Brightness: spectral centroid. Pace: characters per second (spaces included β€” the unit our plan limits use) on a ~15-second passage that every voice of the language read.

Transcription (30 October 2026). 20 recordings per language from FLEURS (native speakers reading Wikipedia-style sentences), transcribed with the same Whisper large-v3-turbo setup the product uses. WER = word error rate, CER = character error rate (the main metric for languages without spaces between words); typ_* = the median clip. Read speech is an easy case β€” real interviews and phone calls produce more errors.

Mixed scripts (15 November 2026). One YouTube-style intro (about 33 words) per language, written (a) in the language's own script, (b) the same with English words such as video, YouTube, subscribe left in English letters, and (c) where people commonly do it, entirely in English letters. Each voice read each version once; Whisper large-v3-turbo transcribed the audio; we counted the language's own words recognised (fuzzy match β‰₯ 0.7) and the English loanwords recognised in either script. no audio = the voice returned no audio after three attempts. One reading per voice and version: treat differences of a few points as noise. Whisper itself is weak in Bengali, Tamil and Telugu (CER 30–40% on FLEURS), so for those languages only the difference between "read" and "skipped" is reliable.

Caveats

  • Measurements describe the voices and model versions available on the dates above; providers update voices.
  • Urdu (Pakistan) voices intermittently returned no audio during the November test β€” a provider-side issue, not a property of the voice.
  • Nothing here is a ranking of "best" voices: pitch and pace are facts, which one fits your content is a choice.

Questions or corrections: audioflowteam@speakscribe.me

Downloads last month
67