Welsh STT v21.03

Welsh STT v21.03 (Dewi Bryn Jones)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by Dewi Bryn Jones and released by the Techiaith Language Technologies Unit
• Model language: Welsh / Cymraeg / cy
• Model date: Accessed from Github on March 31, 2021
• Model type: Speech-to-Text
• Model version: v21.03
• Compatible with 🐸 STT version: v0.9.3
• Code: docker-deepspeech-cy
• License: MIT
• Citation details: @misc{welsh-stt-dewibrynjones, author = {Dewi Bryn Jones}, title = {Docker DeepSpeech Cymraeg}, publisher = {Techiaith}, journal = {docker-deepspeech-cy}, howpublished = {\url{https://github.com/techiaith/docker-deepspeech-cy/releases/tag/21.03}} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Welsh Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

Word Error Rates and Character Error Rates were not reported for this model.

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .76

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

These models were trained with the Welsh dataset from the Common Voice Corpus 6.1 in addition to a small dataset of validated recordings donated by the first users of Bangor University's Language Technology Unit's online automatic transcription website service: Trawsgrifiwr Ar-lein. Detailed release notes here.

Evaluation data

With a language model, the Welsh STT model had a Word Error Rate of 11%. Detailed release notes here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Ukrainian STT v0.4

Ukrainian STT v0.4 (Yurii Paniv)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by Yurii Paniv and released under the voice-recognition-ua project.
• Model language: Ukrainian / українська мова / uk
• Model date: Accessed from Github on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.4
• Compatible with 🐸 STT version: v0.9.3
• Code: voice-recognition-ua
• License: CC BY-NC 4.0
• Citation details: @misc{ukrainian-stt-paniv, author = {Paniv,Yurii}, title = {Ukrainian STT}, publisher = {voice-recognition-ua}, journal = {Github}, howpublished = {\url{https://github.com/robinhad/voice-recognition-ua/releases/tag/v0.4}}, commit={1252a9e9337ceeff52fe9772dc8802f4337ccff3} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Ukrainian Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates (WER) are reported on Github.

Test Corpus	WER	CER
Common Voice	57.2%	16.3%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .76

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model is trained on a total of 1,230 hours from the Ukrainian Dataset at Academic torrents and Common Voice Ukrainian 6.1.

Evaluation data

This model was tested on Common Voice Ukrainian 6.1.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be misused to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Spanish STT v0.0.1

Spanish STT v0.0.1 (Jaco-Assistant)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by DANBER and released under the Jaco-Assistant project.
• Model date: Accessed from Gitlab on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: scribosermo
• License: GNU Lesser General Public License
• Citation details: @misc{spanish-jaco, author = {DANBER}, title = {Spanish Jaco-Assistant}, publisher = {Jaco-Assistant}, journal = {Gitlab}, howpublished = {\url{https://gitlab.com/Jaco-Assistant/Scribosermo}}, commit = {dfc541d2} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Spanish Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates (WER) are reported on Gitlab.

Test Corpus	WER	CER
Common Voice	16.5%	7.6%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: ``

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on the following corpora: CommonVoice + CssTen + LinguaLibre + Mailabs + Tatoeba + Voxforge.

Evaluation data

The model was tested on the Common Voice corpus.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Polish STT v0.0.1

Polish STT v0.0.1 (Jaco-Assistant)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by DANBER and released under the Jaco-Assistant project.
• Model date: Accessed from Gitlab on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: scribosermo
• License: GNU Lesser General Public License
• Citation details: @misc{polish-jaco, author = {DANBER}, title = {Polish Jaco-Assistant}, publisher = {Jaco-Assistant}, journal = {Gitlab}, howpublished = {\url{https://gitlab.com/Jaco-Assistant/Scribosermo}}, commit = {dfc541d2} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Polish Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates and Character Error Rates are reported on Jaco-Assistant.

Test Corpus	WER	CER
Common Voice	3.4%	2.0%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: ``

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on the following corpora: Common Voice + LinguaLibre + Mailabs. Read more about training here.

Evaluation data

The Model was evaluated on the Common Voice corpus. Read more about evaluation here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Komi-Zyrian STT v0.0.1

Komi-Zyrian STT v0.0.1 (ITML)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by Nils Hjortnæs and the Inclusive Technology for Marginalised Languages group.
• Model date: March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• License: AGPL
• Citation details: @inproceedings{hjortnaes-etal-2020-towards, title = "Towards a Speech Recognizer for {K}omi, an Endangered and Low-Resource Uralic Language", author = "Hjortnaes, Nils and Partanen, Niko and Rie{\ss}ler, Michael and M. Tyers, Francis", booktitle = "Proceedings of the Sixth International Workshop on Computational Linguistics of Uralic Languages", month = "1", year = "2020", address = "Wien, Austria", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2020.iwclul-1.5", pages = "31--37" }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Komi-Zyryan Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates and Character Error Rates are reported in the paper.

Test Corpus	WER	CER
Common Voice	70.9%	100%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .95

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on the following corpora: BasFormtask, BasSprecherinnen, Common Voice, CssTen, Gothic, LinguaLibre, Kurzgesagt, Mailabs, MussteWissen, PulsReportage, SWC, Tatoeba, TerraX, Tuda, Voxforge, YKollektiv, ZamiaSpeech, and Common Voice Single Words. Read more about training here.

Evaluation data

The Model was evaluated on Tuda and Common Voice. Read more about evaluation here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Kinyarwanda STT v0.0.1

Kinyarwanda STT v0.0.1 (Digital Umuganda)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally released by Digital Umuganda.
• Model date: Accessed from Github on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: deepspeech-kinyarwanda
• License: MPL 2.0
• Citation details: @misc{deepspeech-kinyarwanda, author = {Digital Umuganda}, title = {Kinyarwanda STT}, publisher = {Digital Umuganda}, journal = {Github}, howpublished = {\url{https://github.com/Digital-Umuganda/Deepspeech-Kinyarwanda}}, commit = {7dbf6705ee38d87138f3558a21f045c40b93f083}}
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Kinyarwanda Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

Test Corpus	WER	CER
Common Voice	60.1%	23.5%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .69

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

Train on approximately 1,200 hours from the Common Voice corpus.

Evaluation data

Evaluated on the test set from the Common Voice corpus.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

Italian STT v0.0.1

Italian STT v0.0.1 (Jaco-Assistant)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by DANBER and released under the Jaco-Assistant project.
• Model date: Accessed from Gitlab on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: scribosermo
• License: GNU Lesser General Public License
• Citation details: @misc{italian-jaco, author = {DANBER}, title = {Italian Jaco-Assistant}, publisher = {Jaco-Assistant}, journal = {Gitlab}, howpublished = {\url{https://gitlab.com/Jaco-Assistant/Scribosermo}}, commit = {dfc541d2} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Italian Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates (WER) are reported on Gitlab.

Test Corpus	WER	CER
Common Voice	24.9%	9.4%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: ``

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on approximately ~257 hours from the following corpora: Common Voice + LinguaLibre + Mailabs + Voxforge. Read more about training here.

Evaluation data

The model was tested on approximately ~21 hours from Common Voice. Read more about testing here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

German STT v0.0.1

German STT v0.0.1 (Jaco-Assistant)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by DANBER and released under the Jaco-Assistant project.
• Model date: Accessed from Gitlab on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: scribosermo
• License: GNU Lesser General Public License
• Citation details: @misc{german-jaco, author = {DANBER}, title = {German STT for Jaco-Assistant}, publisher = {Jaco-Assistant}, journal = {Gitlab}, howpublished = {\url{https://gitlab.com/Jaco-Assistant/Scribosermo}}, commit = {dfc541d2} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the German Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates and Character Error Rates are reported on Jaco-Assistant.

Test Corpus	WER	CER
Common Voice	12.8%	5.6%
Tuda	24.6%	10.1%
CommonVoice + Tuda + Voxforge	16.2%	6.9%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: ``

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on the following corpora: BasFormtask, BasSprecherinnen, Common Voice, CssTen, Gothic, LinguaLibre, Kurzgesagt, Mailabs, MussteWissen, PulsReportage, SWC, Tatoeba, TerraX, Tuda, Voxforge, YKollektiv, ZamiaSpeech, and Common Voice Single Words. Read more about training here.

Evaluation data

The Model was evaluated on Tuda and Common Voice. Read more about evaluation here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

German STT v0.9.0

German STT v0.9.0 (Aashish Agarwal)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by Aashish Agarwal and released under the deepspeech-german project.
• Model date: Accessed from deepspeech-german on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.9.0
• Compatible with 🐸 STT version: v0.9.3
• Code: deepspeech-german
• License: Apache 2.0
• Citation details: @inproceedings{agarwal-zesch-2019-german, author = "Aashish Agarwal and Torsten Zesch", title = "German End-to-end Speech Recognition based on DeepSpeech", booktitle = "Preliminary proceedings of the 15th Conference on Natural Language Processing (KONVENS 2019): Long Papers", year = "2019", address = "Erlangen, Germany", publisher = "German Society for Computational Linguistics \& Language Technology", pages = "111--119" }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the German Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

No exact statistics on transcription accuracy, however, Word Error Rate was in the range of 10% to 20% on Mozilla and Tuda-De test set. Relevant discussion here.

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .69

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This German STT model was bootstrapped from a pre-trained English model, and fine-tuned to German via the following datasets: Common Voice 5.1 (750 hours) + SWC (248 hours) + MAILABS (233 hours) + Tuda-De (184 hours) + Voxforge (57 hours).

Evaluation data

This German STT model was evaluated on the following datasets: Common Voice 5.1 and Tuda-De.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.

French STT v0.0.1

French STT v0.0.1 (Jaco-Assistant)

Jump to section:

• Model details
• Intended use
• Performance Factors
• Metrics
• Training data
• Evaluation data
• Ethical considerations
• Caveats and recommendations

Model details

• Person or organization developing model: Originally trained by DANBER and released under the Jaco-Assistant project.
• Model date: Accessed from Gitlab on March 31, 2021
• Model type: Speech-to-Text
• Model version: v0.0.1
• Compatible with 🐸 STT version: v0.9.3
• Code: scribosermo
• License: GNU Lesser General Public License
• Citation details: @misc{french-jaco, author = {DANBER}, title = {French DeepSpeech for Jaco-Assistant}, publisher = {Jaco-Assistant}, journal = {Gitlab}, howpublished = {\url{https://gitlab.com/Jaco-Assistant/Scribosermo}}, commit = {dfc541d2} }
• Where to send questions or comments about the model: You can leave an issue on STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the French Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

The following Word Error Rates (WER) are reported on Gitlab.

Test Corpus	WER	CER
Common Voice	19.5%	9.2%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: ``

Model Size

model.pbmm: 181M
model.tflite: 46M

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This French STT model was trained on approximately 787 hours of Common Voice + CssTen + LinguaLibre + Mailabs + Tatoeba + Voxforge. Read more about training here.

Evaluation data

This French STT model was tested on approximately 25 hours of Common Voice. Read more about testing here.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.