Persian STT v0.1.0

Persian STT v0.1.0

Jump to section:

• Model card for Persian STT v0.1.0
  • Model details
  • Intended use
  • Performance Factors
  • Metrics
    • Transcription Accuracy
    • Real-Time Factor
    • Model Size
    • Approaches to uncertainty and variability
  • Training data
  • Evaluation data
  • Ethical considerations
    • Demographic Bias
    • Surveillance
  • Caveats and recommendations

Model details

• Person or organization developing model: Maintained by oct4pie.
• Model language: Persian / Farsi / fa, fa-IR
• Model date: June 21, 2022
• Model type: Speech-to-Text
• Model version: v0.1.0
• Compatible with 🐸 STT version: v1.3.0
• License: GNU Lesser General Public License v3.0
• Citation details: @techreport{persian-stt, author = {Mehdi Hajmollaahmad Naraghi}, title = {Persian STT v0.1.0}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2022}, month = {June}, number = {STT-FA-0.1.0} }
• persian-tts GitHub Repo
• Where to send questions or comments about the model: You can leave an issue on STT issues, open a new discussion on STT discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Persian 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

Using the language model with settings lm_alpha=0.36669178512950323 and lm_beta=0.3457913671678824 (found via lm_optimizer.py):

• Common-Voice clean: WER: 10.81%, CER: 2.506%
• More about the model at persian-tts repo

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: .65

Model Size

For STT, you always must deploy an acoustic model, and it is often the case you also will want to deploy an application-specific language model.

Model type	Filename	Size
Acoustic model (tflite)	model.tflite	45.3M
Language model	kenlm.scorer	1.63GB

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 9.0 Persian (cleaned and with custom train/dev/test splits). In total approximately ~271 hours of data.

Evaluation data

The validation ("dev") sets were cleaned and generated from Common Voice 9.0 Persian.

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.

Hindi STT v0.8.99

Hindi STT v0.8.99

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: Trained and released by Bülent Özden a member of Common Voice Türkçe for the 3D Voice Chess project by Harikalar Kutusu.
• Model language: Hindi / हिन्दी / hi
• Model date: March 13, 2022
• Model type: Speech-to-Text
• Model version: v0.8.99
• Compatible with 🐸 STT version: v1.0.0
• License: CC-BY-SA 4.0
• Citation details: @misc{hindi-stt, author = {Bülent Özden}, title = {Hindi STT v0.8.99}, institution = {Harikalar Kutusu}, address = {\url{https://coqui.ai/models}} year = {2022}, month = {March}, number = {STT-HI-0.8.99} }
• Where to send questions or comments about the model: You can leave an issue on STT issues, open a new discussion on STT discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Hindi 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

This model only includes acoustic model, as it is developed for the special purpose low-vocabulary application. The following is the results from the Acoustic Model training.

Test Corpus	WER	CER
Common Voice	82.2%	34.6%

Model Size

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 Corpus 8.0 for Hindi. Custom train/dev/test splits are created by Common Voice Corpora Creator with --duplicate-sentence-count 99 parameter, which allowed us to use the whole dataset. The dataset contains approximately ~11 hours of voice data (276 distinct voices, 65% male, 4% female).

Note: Our model numbering for Common Voice only data reflect Common Voice corpus version and Corpora Creator duplicate-sentence-count (dsc) setting (e.g. "v0.corpus.dsc").

Evaluation data

The validation ("dev") and test ("test") sets also came from CV as specified above.

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 v0.9

French STT v0.9

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 and released by the commonvoice-fr project, revived by Waser Technologies
• Model date: Accessed from Github on Jun 10, 2022
• Model type: Speech-to-Text
• Model version: v0.9
• Compatible with 🐸 STT version: v1.4.0
• Code: commonvoice-fr
• License: MPL 2.0
• Citation details: @misc{commonvoice-fr, author = {commonvoice-fr Contributors}, title = {Common Voice Fr STT Model}, publisher = {Github}, journal = {GitHub repository}, howpublished = {\url{https://github.com/wasertech/commonvoice-fr/releases/tag/v0.9.0-fr-0.1}}, commit = {0a2d028b124691bbee656f43aa02251169dce69b} }
• 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 Github.

Test Corpus	WER	CER
African_Accented_French_test.csv	47.7%	6.6%
Att-HACK	12.9%	7.1%
M-AILABS	9.9%	3.3%
trainingspeech	10.9%	4.1%
Common Voice	31.5%	15.2%
LinguaLibre	67.6%	21.6%
MLS	22.6%	9.7%

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: ~0.3

Model Size

model.tflite: 46M
kenlm.scorer: 689M

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 the following corpora:

1. Lingua Libre (~40h)
2. Common Voice FR (v8) (~850h, by allowing up to 32 duplicates)
3. Training Speech (~180h)
4. African Accented French (~15h)
5. M-AILABS French (~315h)
6. Multilingual LibriSpeech (~1,100h)
7. Att-HACK (~75h)

Total : ~2,573h
(~1,925h by default)

Evaluation data

The model was tested on the following corpora.

1. Lingua Libre
2. Common Voice FR (v9)
3. Training Speech
4. African Accented French
5. M-AILABS French
6. Multilingual LibriSpeech
7. Att-HACK

Data was augmented with the following parameters.

Parsed augmentations: [
    Reverb(p=0.1, delay=ValueRange(start=50.0, end=50.0, r=30.0), decay=ValueRange(start=10.0, end=2.0, r=1.0)),
    Resample(p=0.1, rate=ValueRange(start=12000, end=8000, r=4000)),
    Codec(p=0.1, bitrate=ValueRange(start=48000, end=16000, r=0)),
    Volume(p=0.1, dbfs=ValueRange(start=-10.0, end=-40.0, r=0.0)),
    Pitch(p=0.1, pitch=ValueRange(start=1.0, end=1.0, r=0.2)),
    Tempo(p=0.1, factor=ValueRange(start=1.0, end=1.0, r=0.5), max_time=-1.0), 
    FrequencyMask(p=0.1, n=ValueRange(start=1, end=3, r=0), size=ValueRange(start=1, end=5, r=0)), 
    TimeMask(p=0.1, domain='signal', n=ValueRange(start=3, end=10, r=2), size=ValueRange(start=50.0, end=100.0, r=40.0)),
    Dropout(p=0.1, domain='spectrogram', rate=ValueRange(start=0.05, end=0.05, r=0.0)),
    Add(p=0.1, domain='signal', stddev=ValueRange(start=0.0, end=0.0, r=0.5)),
    Multiply(p=0.1, domain='features', stddev=ValueRange(start=0.0, end=0.0, r=0.5))
]

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.

Czech STT v0.3.0

Czech STT v0.3.0

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: Trained by Vojtěch Drábek.
• Model language: Czech / čeština / cs
• Model date: May 31, 2022
• Model type: Speech-to-Text
• Model version: v0.3.0
• Compatible with 🐸 STT version: v0.9.3
• License: CC-BY-NC 4.0
• Citation details: @misc{czech-stt, author = {Drábek, Vojtěch}, title = {Czech STT 0.3}, publisher = {comodoro}, journal = {deepspeech-cs}, howpublished = {\url{https://github.com/comodoro/deepspeech-cs}} }
• Where to send questions or comments about the model: You can leave an issue on the model release page or STT-model issues, open a new discussion on STT-model discussions, or chat with us on Gitter or Matrix channel coqui-ai/STT.

Intended use

Speech-to-Text for the Czech 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

More information reported on Github.

Test Corpus	WER	CER
Acoustic model
Czech Common voice 6.1	40.6%	10.7%
Vystadial 2016	50.6%	19.6%
Parliament Plenary Hearings	21.3%	5.3%
ParCzech 3.0	21%	6.2%
With the attached scorer
Czech Common voice 6.1	15.3%	6.8%
Vystadial 2016	35.7%	20.1%
Parliament Plenary Hearings	9.7%	3.7%
ParCzech 3.0	10.1%	4.5%

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: 0.73

Model Size

model.tflite: 46M
scorer: 461M

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:

1. Vystadial 2016 – Czech data
2. OVM – Otázky Václava Moravce
3. Czech Parliament Meetings
4. Large Corpus of Czech Parliament Plenary Hearings
5. Common Voice Czech
6. Some private recordings and parts of audiobooks

Evaluation data

The model was evaluated on Common Voice Czech, Large Corpus of Czech Parliament Plenary Hearings, Vystadial 2016 – Czech data and ParCzech 3.0 test sets.

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 many 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.

Yoloxóchitl Mixtec STT

Yoloxóchitl Mixtec STT

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 Joe Meyer.
• Model language: Yoloxóchitl Mixtec / / xty
• Model date: April 17, 2022
• Model type: Speech-to-Text
• Model version: v0.1.0
• Compatible with 🐸 STT version: v1.0.0
• License: CC BY-NC-SA 3.0
• Citation details: @techreport{xty-stt, author = {Meyer,Joe}, title = {Yoloxóchitl Mixtec STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2022}, month = {April}, number = {STT-SLR89-XTY-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 Yoloxóchitl Mixtec 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 for a modified data set from OpenSLR SLR89. The official validated.tsv had rows removed which had errors processing, and the data was re-processed by Cmmon Voice Utils to convert to 16kHz mono-channel PCM .wav files.

Test Corpus	WER	CER
OpenSLR	48.85%	18.04%

Model Size

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 a modified data set from OpenSLR SLR89. The official validated.tsv had rows removed which had errors processing, and the data was re-processed by Cmmon Voice Utils to convert to 16kHz mono-channel PCM .wav files.

Evaluation data

This model was evaluated on a modified data set from OpenSLR SLR89. The official validated.tsv had rows removed which had errors processing, and the data was re-processed by Cmmon Voice Utils to convert to 16kHz mono-channel PCM .wav files.

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.

Sierra Totonac STT

Sierra Totonac STT

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 Bülent Özden, a member of Common Voice Türkçe.
• Model language: Totonac / Sierra Totonac / tos
• Model date: April 12, 2022
• Model type: Speech-to-Text
• Model version: v1.0.0
• Compatible with 🐸 STT version: v1.3.0
• License: CC BY-NC-SA 3.0
• Citation details: @techreport{totonac-stt, author = {Bülent Özden}, title = {Totonac STT 1.0}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2022}, month = {April}, number = {STT-TOS-1.0} }
• 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 Sierra Totonac 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
OpenSLR 107	87.5%	25.8%

Model Size

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 Totonac Speech with transcription corpus.

Evaluation data

This model was evaluated on Totonac Speech with transcription 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.

Western Highland Chatino STT

Western Highland Chatino STT

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 Bülent Özden, a member of Common Voice Türkçe.
• Model language: Western Highland Chatino / ctp
• Model date: 12th April, 2022
• Model type: Speech-to-Text
• Model version: v1.0.0
• Compatible with STT version: v1.3.0
• License: CC-BY-SA 4.0
• 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 Western Highland Chatino 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
GORILLA	77.2%	30.9%

Model Size

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 GORILLA ctp

Citation

• Malgorzata E. Cavar, Damir Cavar, Hilaria Cruz (2016) "Endangered Language Documentation: Bootstrapping a Chatino Speech Corpus, Forced Aligner, ASR". Pages 4004-4011 Of N. Calzolari (et al. eds) Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016) in Portorož, Slovenia, European Language Resources Association (ELRA), Paris, France.

Evaluation data

This model was evaluated on GORILLA ctp

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 analyse 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.

Swahili STT v0.8

Swahili STT v8.0 (Coqui)

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: Maintained by Coqui.
• Model language: Swahili / kiswahili / sw
• Model date: March 8, 2022
• Model type: Speech-to-Text
• Model version: v8.0
• Compatible with 🐸 STT version: v1.3.0
• License: Apache 2.0
• Citation details: @techreport{swahili-stt, author = {Coqui}, title = {Swahili STT v8.0}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2022}, month = {March}, number = {STT-SW-8.0} }
• Where to send questions or comments about the model: You can leave an issue on STT issues, open a new discussion on STT discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Swahili 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

Using the language model with settings lm_alpha=0.898202045251655 and lm_beta=2.2684674938753755 (found via lm_optimizer.py):

• Swahili Common Voice 8.0 Test: WER: 15.8%, CER: 6.6%

Model Size

For STT, you always must deploy an acoustic model, and it is often the case you also will want to deploy an application-specific language model.

Model type	Vocabulary	Filename	Size
Acoustic model	open	model.tflite	45M
Language model	large	large-vocabulary.scorer	321M

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 8.0 Swahili.

Evaluation data

The validation ("dev") sets came from Common Voice 8.0.

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 v0.8

French STT v0.8 (commonvoice-fr)

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 and released by the commonvoice-fr project, revived by Waser Technologies
• Model date: Accessed from Github on February 9, 2022
• Model type: Speech-to-Text
• Model version: v0.8
• Compatible with 🐸 STT version: v1.2.0
• Code: commonvoice-fr
• License: MPL 2.0
• Citation details: @misc{commonvoice-fr, author = {commonvoice-fr Contributors}, title = {Common Voice STT Model}, publisher = {Github}, journal = {GitHub repository}, howpublished = {\url{https://github.com/wasertech/commonvoice-fr/releases/tag/v0.8.0-fr-0.3}}, commit = {0a2d028b124691bbee656f43aa02251169dce69b} }
• 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 Github.

Test Corpus	WER	CER
African_Accented_French_test.csv	43.6%	24.8%
Att-HACK	12.8%	6.0%
M-AILABS	12.2%	3.7%
trainingspeech	12.1%	4.0%
Common Voice	37.0%	19.4%
LinguaLibre	59.3%	21.3%
MLS	26.8%	12.2%

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.tflite: 46M
kenlm.scorer: 689M

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 the following corpora:

1. Lingua Libre (~40h)
2. Common Voice FR (v8) (~826h, by allowing up to 32 duplicates)
3. Training Speech (~180h)
4. African Accented French (~15h)
5. M-AILABS French (~315h)
6. Multilingual LibriSpeech (~1,100h)
7. Att-HACK (~75h)

Total : ~2,551h
(~1,903h by default)

Evaluation data

The model was tested on the following corpora.

1. Lingua Libre
2. Common Voice FR (v8)
3. Training Speech
4. African Accented French
5. M-AILABS French
6. Multilingual LibriSpeech
7. Att-HACK

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.

English STT v1.0.0-huge-vocab

English STT v1.0.0 (Huge Vocabulary)

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: Maintained by Coqui.
• Model language: English / English / en
• Model date: October 3, 2021
• Model type: Speech-to-Text
• Model version: v1.0.0
• Compatible with 🐸 STT version: v1.0.0
• License: Apache 2.0
• Citation details: @techreport{english-stt, author = {Coqui}, title = {English STT v1.0.0}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2021}, month = {October}, number = {STT-EN-1.0.0} }
• Where to send questions or comments about the model: You can leave an issue on STT issues, open a new discussion on STT discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the English 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

Using the huge-vocabulary.scorer language model:

• Librispeech clean: WER: 4.5%, CER: 1.6%
• Librispeech other: WER: 13.6%, CER: 6.4%

Model Size

For STT, you always must deploy an acoustic model, and it is often the case you also will want to deploy an application-specific language model.

Model type	Vocabulary	Filename	Size
Acoustic model	open	model.tflite	181M
Language model	large	huge-vocabulary.scorer	923M

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 7.0 English (custom Coqui train/dev/test splits), LibriSpeech, and Multilingual Librispeech. In total approximately ~47,000 hours of data.

Evaluation data

The validation ("dev") sets came from CV, Librispeech, and MLS.

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.