Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
aac
unknown
__key__
stringlengths
26
31
__url__
stringclasses
11 values
[ 255, 241, 76, 128, 40, 31, 252, 222, 2, 0, 76, 97, 118, 99, 54, 48, 46, 51, 49, 46, 49, 48, 50, 0, 66, 85, 28, 61, 255, 255, 224, 2, 82, 197, 83, 180, 48, 245, 77, 231, 28, 115, 123, 215, 10, 145, 50, 93, 41, 46, 213...
it_xectz_20250729_1500_seg133
hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar.gz
"//FMgCef/N4CAExhdmM2MC4zMS4xMDIAQjUZH///YAIuw1Sw0VhwLQsHUM13nXjiPOyVOKy1aIy7x90lcXW9Dw2zxeIvNntcZ2u(...TRUNCATED)
es_xectz_20250729_1500_seg77
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCWf/N4CAExhdmM2MC4zMS4xMDIAQlUd+///8AJW2VBiapykb1WXZmppKpdXV02fYgQf1O1Ijxe2nY+e48O0xmOvGma0xGE(...TRUNCATED)
es_xectz_20250729_1500_seg60
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCef/N4CAExhdmM2MC4zMS4xMDIAQjUd3///8AJCy1WBsPRsHTK31xvnzErXbi5kEkVqlPWkXV5BFcnk+n7/kKkPm5EsO0A(...TRUNCATED)
ca_xectz_20250729_1500_seg119
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCVf/N4CAExhdmM2MC4zMS4xMDIAQjUf////gAI+xVzWJOt8b3574i8qdZiSxdR7ZxUSprnge5/EeN837fAYWmlSePNtw3T(...TRUNCATED)
ca_xectz_20250729_1500_seg113
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCZ//N4CAExhdmM2MC4zMS4xMDIAQlUf3f//8AJK4U1h6puUrvjprLJvVmcZqrUm3yggH/D4UmAPh6ft/C/g9h/R/H9N7PT(...TRUNCATED)
it_xectz_20250729_1500_seg122
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCd//N4CAExhdmM2MC4zMS4xMDIAQlUT1//+UAIuyVKxQPRQHQsLUv0uuplpdSmsjFy4QC2JIlih8/09lvCuSNOwn9vnXa6(...TRUNCATED)
en_xectz_20250729_1434_seg146
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCS//N4CAExhdmM2MC4zMS4xMDIAQjUWi///wAI6w13WN+29ZSVxqmXmk3wq6uKfd3lyomWPqsrl/uQsL3/EM3FjUR7Q9Yr(...TRUNCATED)
ca_xectz_20250729_1500_seg50
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCU//N4CAExhdmM2MC4zMS4xMDIAQjUe////YAJSw1vQwPQsHUOOtxqqee2r71eXlcIR7SpqrqFWHPHxn5eAwr3xuAcNHY+(...TRUNCATED)
it_xectz_20250729_1500_seg71
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
"//FMgCc//N4CAExhdmM2MC4zMS4xMDIAQlUbX///0AJGw1aywLQsHVO+uN3zqxWqyLtW9TNQe6Stb1xruD7p6T+jmYeoYzKEFAb(...TRUNCATED)
pt_xectz_20250729_1500_seg107
"hf://datasets/tachiwin/tachiwin_voice_raw@c376ce4118ddbf43e331912845a59d41fb3ed7c5/other_250729.tar(...TRUNCATED)
End of preview. Expand in Data Studio

Tachiwin Indigenous Languages of Mexico Voice Pretrain Dataset (RAW)

This dataset collects a large amount of speech hours in indigenous languages of Mexico recorded from 23 public radio stations that broadcast to the indigenous communities.

How it was obtained

After the raw recording of all transmissions, the music and inaudible sound were removed and then the samples of pure speech were classified by language. However, as there are no existing speech language classifiers for the indigenous languages of Mexico (the creation of these is the aim of this dataset), we classified into two groups:

  • Mainstream (English, Spanish, and French samples), as they are known to be present on the broadcasts.
  • Other languages, which almost certainly correlate with indigenous languages, as other mainstream languages are extremely rare on these broadcasts.

In this raw dataset only the other languages are included and therefore the prefix depicting the identified language must be ignored. Nonetheless, we found some consistent correlation between identified and actual languages (e.g., "Japanese" to the Zapotec superfamily). These could be used in post-processing steps as hints to identify the actual languages. Additionally, knowing the geographical location of the broadcasters and the name of the station, we devised probable languages found on each transmission. For example, audios with prefix xectz from XECTZ. La Voz de la Sierra Norte contain speech in Highland Totonac (tos) and Northern Puebla Nahuatl (nhe), along with Spanish and, more rarely, English and French. See the table below.

Aims

The aim of this dataset is to pretrain (without language identification) language models for text recognition or speech, or to be annotated by the given hints for identification and to fine-tune or train models.

Format

Recordings are tar.gz compressed by day and are .aac 22 kHz recorded directly from the public radio broadcast. Some stations broadcast in lower quality, so perceptual quality may be worse. Further quality classification may be needed.

Disclaimer

  • Although music was stripped, some background music with copyright may be found in very short segments.
  • As these are public people-owned radios and some stations allow external ads, it is possible that a few segments contain advertisements with unethical claims. Please report if you find such content.
  • Political speeches from government, parties, and journalists may also appear, as this is very common in the content.

All content collected, preprocessed, and made available by Tachiwin is open-source, with the mission to diminish the technological gap for the indigenous languages of Mexico and the world, making high-end technology available to the communities — always free and always sovereign.


Probable Indigenous Languages per Station

Station Name Station Code Probable Indigenous Language Codes
La Voz de las Huastecas xeant hus, nhw
La Voz de los Vientos xecopa zos, tzh, tzo
La Voz de la Sierra Norte xectz azz, tos, otm
La Voz de los Tres Ríos xeetch mfy, yaq
La Voz de la Sierra Juárez xeglo zaa, chq, cco
La Voz de la Costa Chica xejam mxt, miy, amu, azm, azg, clo, chd
La Voz de los Cuatro Pueblos xejmn crn, hch
La Voz de la Chinantla xeojn chj, cuc, cnt, cle, chz, ctl, cpa, cte, cso, chq, cco, cnl
La Voz de los Mayas xepet yua
La Voz de los P’urhepechas xepur pua, tsz
La Voz del Valle xeqin dih, klb, ppi
La Voz de la Sierra Tarahumara xetar tar, rar, ntp, tla
La Voz de la Mixteca xetla mie, mim, miy, mix, mjc, miz, mdv, mil, mig, mio, mks, mxb, mxy, meh, xta, xtd, xti, xtj, xtm, xtn, mbz, vmm, vmj, mip, mit
Las Tres Voces de Durango xetph azn, azd, ntp
La Voz de la Frontera Sur xevfs toj, tzh, tzo, ctp, knj, quc, kic, jac, ixq, mam, kek
La Voz del Corazón de la Selva xexpuj yua, ctp, cac, jac, kic, mam, kek
La Voz de la Sierra de Zongolica xezon nlv, maz
La Voz de la Montaña xezv ngu, tcf, tpl, tpc, tpx
La Voz del Pueblo Hñähñú xhcarh ote, otq
La Voz de los Chontales xhcpbs chf
La Voz del Rio Yaqui xhcpct yaq
La Voz del Gran Pueblo xenka yua
La Voz de la Sierra Oriente xetumi mmc, tsz

License

This model is released under the MIT License. See the LICENSE file for details.

Contributing

We welcome contributions from indigenous language communities and researchers:

  • Report issues with specific language variants
  • Contribute additional training data
  • Suggest improvements for regional variants
  • Share use cases and applications

Contact

For questions, collaborations, or community partnerships, please open an issue in this repository.


In Memoriam Fidencio Hernández (1990-2025) founder of this initiative. May the indigenous languages of Mexico and the world are never lost. Tachiwin is an initiative commited to the preservation of indigenous languages of Mexico.


Dataset en Lenguas Originarias de México para el Pre-Entrenamiento de Modelos de Voz - Español

Este conjunto de datos recopila una gran cantidad de horas de habla en lenguas indígenas de México, grabadas de 23 estaciones de radio públicas que transmiten hacia las comunidades indígenas.

Cómo se obtuvo

Tras la grabación en bruto de todas las transmisiones, se eliminaron la música y los sonidos inaudibles, y luego las muestras de habla pura fueron clasificadas por idioma. Sin embargo, dado que no existen clasificadores de lenguas indígenas de México (la creación de estos es el objetivo de este dataset), se clasificaron en dos grupos:

  • Lenguas dominantes (inglés, español y francés), ya que se sabe que están presentes en las transmisiones.
  • Otras lenguas, que casi con certeza corresponden a lenguas indígenas, pues la presencia de otras lenguas dominantes es extremadamente rara en estas transmisiones.

En este dataset en bruto solo se incluyen las otras lenguas, por lo que el prefijo que representa el idioma identificado debe ignorarse. Aun así, encontramos correlaciones consistentes entre lenguas identificadas y lenguas reales (por ejemplo, "japonés" con la superfamilia zapoteca). Estas pueden usarse en pasos posteriores de procesamiento como pistas para identificar las lenguas reales. Además, conociendo la ubicación geográfica de las emisoras y el nombre de la estación, dedujimos las lenguas probables en cada transmisión. Por ejemplo, los audios con prefijo xectz de XECTZ. La Voz de la Sierra Norte contienen habla en totonaco de la Sierra (tos) y náhuatl del norte de Puebla (nhe), junto con español y, más raramente, inglés y francés. Véase la tabla a continuación.

Objetivos

El objetivo de este conjunto de datos es el preentrenamiento (sin identificación de lengua) de modelos de lenguaje para reconocimiento de texto o habla, o para ser anotados con las pistas proporcionadas para identificación y posteriormente usados en entrenamiento o ajuste fino.

Formato

Las grabaciones están comprimidas por día en archivos tar.gz y son .aac a 22 kHz grabadas directamente de la transmisión de radio pública. Algunas emisoras transmiten en menor calidad, por lo que la calidad perceptual puede ser inferior. Puede ser necesario un posterior filtrado por calidad.

Aviso

  • Aunque se eliminó la música, pueden encontrarse fragmentos muy cortos con música de fondo protegida por derechos de autor.
  • Al ser radios públicas y comunitarias, y dado que algunas permiten anuncios externos, es posible que ciertos segmentos contengan publicidad con afirmaciones poco éticas. Por favor repórtelo si encuentra alguno.
  • También pueden encontrarse discursos políticos de gobierno, partidos o periodistas, ya que son frecuentes en este tipo de contenidos.

Todo el contenido recopilado, preprocesado y puesto a disposición por Tachiwin es de código abierto, con la misión de reducir la brecha tecnológica para las lenguas originarias de México y del mundo, poniendo tecnología de alto nivel al alcance de las comunidades — siempre libre y siempre soberano.


Lenguas indígenas probables por estación

Nombre de la estación Código Códigos de lenguas indígenas probables
La Voz de las Huastecas xeant hus, nhw
La Voz de los Vientos xecopa zos, tzh, tzo
La Voz de la Sierra Norte xectz azz, tos, otm
La Voz de los Tres Ríos xeetch mfy, yaq
La Voz de la Sierra Juárez xeglo zaa, chq, cco
La Voz de la Costa Chica xejam mxt, miy, amu, azm, azg, clo, chd
La Voz de los Cuatro Pueblos xejmn crn, hch
La Voz de la Chinantla xeojn chj, cuc, cnt, cle, chz, ctl, cpa, cte, cso, chq, cco, cnl
La Voz de los Mayas xepet yua
La Voz de los P’urhepechas xepur pua, tsz
La Voz del Valle xeqin dih, klb, ppi
La Voz de la Sierra Tarahumara xetar tar, rar, ntp, tla
La Voz de la Mixteca xetla mie, mim, miy, mix, mjc, miz, mdv, mil, mig, mio, mks, mxb, mxy, meh, xta, xtd, xti, xtj, xtm, xtn, mbz, vmm, vmj, mip, mit
Las Tres Voces de Durango xetph azn, azd, ntp
La Voz de la Frontera Sur xevfs toj, tzh, tzo, ctp, knj, quc, kic, jac, ixq, mam, kek
La Voz del Corazón de la Selva xexpuj yua, ctp, cac, jac, kic, mam, kek
La Voz de la Sierra de Zongolica xezon nlv, maz
La Voz de la Montaña xezv ngu, tcf, tpl, tpc, tpx
La Voz del Pueblo Hñähñú xhcarh ote, otq
La Voz de los Chontales xhcpbs chf
La Voz del Rio Yaqui xhcpct yaq
La Voz del Gran Pueblo xenka yua
La Voz de la Sierra Oriente xetumi mmc, tsz

Licencia

Este modelo se publica bajo la Licencia MIT. Consulte el archivo LICENSE para más detalles.

Contribuciones

Agradecemos las contribuciones de las comunidades de lenguas indígenas y los investigadores:

  • Reporte problemas con variantes lingüísticas específicas
  • Contribuya con datos de entrenamiento adicionales
  • Sugiera mejoras para variantes regionales
  • Comparta casos de uso y aplicaciones

Contacto

Para preguntas, colaboraciones o alianzas comunitarias, por favor abra un issue en este repositorio.


En Memoria de Fidencio Hernández (1990-2025) fundador de esta iniciativa. Que las lenguas indígenas de México y del mundo nunca se pierdan.

Tachiwin es una iniciativa comprometida con la preservación de las lenguas indígenas de México.


Downloads last month
458