Skip to content

Releases: coqui-ai/STT-models

Wolof STT v0.1.0

29 Apr 16:22
Compare
Choose a tag to compare

Wolof STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Wolof / Wolof / wo
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{wolof-stt, author = {Tyers,Francis}, title = {Wolof STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-ALFFA-WO-0.1} }
  • 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 Wolof 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 omnilingo.

Test Corpus WER CER
ALFFA 57.1% 18.1%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Wolof subset of the ALFFA corpus.

Evaluation data

The Model was evaluated on the Wolof subset of the ALFFA 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.

Romansh Sursilvan STT v0.1.1

29 Apr 15:38
Compare
Choose a tag to compare

Romansh Sursilvan STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Romansh Sursilvan / romontsch sursilvan / rm-sursilvan
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{rm-sursilvan-stt, author = {Tyers,Francis}, title = {Romansh Sursilvan STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-RM_SURSILVAN-0.1} }
  • 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 Romansh Sursilvan 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 omnilingo.

Test Corpus WER CER
Common Voice 79.6% 23.9%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Amharic STT v0.1.0

29 Apr 16:08
Compare
Choose a tag to compare

Amharic STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Amharic / አማርኛ / am
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{amharic-stt, author = {Tyers,Francis}, title = {Amharic STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-ALFFA-AM-0.1} }
  • 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 Amharic 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 omnilingo.

Test Corpus WER CER
ALFFA 75.1% 29.4%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Amharic subset of the ALFFA corpus.

Evaluation data

The Model was evaluated on the Amharic subset of the ALFFA 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.

Upper Sorbian STT v0.1.0

29 Apr 08:56
Compare
Choose a tag to compare

Upper Sorbian STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Upper Sorbian / Hornjoserbšćina / hsb
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{upper-sorbian-stt, author = {Tyers,Francis}, title = {Upper Sorbian STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-HSB-0.1} }
  • 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 Upper Sorbian 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 omnilingo.

Test Corpus WER CER
Common Voice 92.3% 32.4%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Turkish STT v0.1.0

28 Apr 14:12
Compare
Choose a tag to compare

Turkish STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Turkish / Türkçe / tr
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{turkish-stt, author = {Tyers,Francis}, title = {Turkish STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-TR-0.1} }
  • 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 Turkish 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 omnilingo.

Test Corpus WER CER
Common Voice 89.3% 30.8%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Thai STT v0.1.0

28 Apr 13:53
Compare
Choose a tag to compare

Thai STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Thai / ภาษาไทย / th
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{thai-stt, author = {Tyers,Francis}, title = {Thai STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-TH-0.1} }
  • 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 Thai 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 omnilingo.

Test Corpus WER CER
Common Voice 100% 36.0%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Tatar STT v0.1.0

28 Apr 13:46
Compare
Choose a tag to compare

Tatar STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Tatar / Татарча / tt
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{tatar-stt, author = {Tyers,Francis}, title = {Tatar STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-TT-0.1} }
  • 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 Tatar 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 omnilingo.

Test Corpus WER CER
Common Voice 85.8% 31.7%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Romansh Vallader STT v0.1.0

28 Apr 15:08
Compare
Choose a tag to compare

Romansh Vallader STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Romansh Vallader / Rumantsch / rm-vallader
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{rm-vallader-stt, author = {Tyers,Francis}, title = {Romansh Vallader STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-RM_VALLADER-0.1} }
  • 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 Romansh Vallader 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 omnilingo.

Test Corpus WER CER
Common Voice 84.0% 26.2%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Basque STT v0.1.1

28 Apr 14:49
Compare
Choose a tag to compare

Basque STT v0.1.1 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Basque / Euskara / eu
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.1
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{basque-stt, author = {Tyers,Francis}, title = {Basque STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-EU-0.1} }
  • 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 Basque 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 omnilingo.

Test Corpus WER CER
Common Voice 68.7% 15.6%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.

Tamil STT v0.1.0

27 Apr 15:27
Compare
Choose a tag to compare

Tamil STT v0.1.0 (ITML)

Jump to section:

Model details

  • Person or organization developing model: Originally trained by Francis Tyers and the Inclusive Technology for Marginalised Languages group.
  • Model language: Tamil / தமிழ் / ta
  • Model date: April 26, 2021
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v0.9.3
  • License: AGPL
  • Citation details: @techreport{tamil-stt, author = {Tyers,Francis}, title = {Tamil STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-CV6.1-TA-0.1} }
  • 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 Tamil 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 omnilingo.

Test Corpus WER CER
Common Voice 99.9% 46.6%

Real-Time Factor

Real-Time Factor (RTF) is defined as processing-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 Common Voice 6.1 train.

Evaluation data

The Model was evaluated on Common Voice 6.1 test.

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.