Releases: coqui-ai/STT-models
English STT v1.0.0-yesno
English STT v1.0.0 (yesno)
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:
Small vocabulary Speech-to-Text - Model version:
yesno-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
STTissues, open a new discussion onSTTdiscussions, or chat with us on Gitter.
Intended use
Closed vocabulary ("yes" and "no") Speech-to-Text for the English Language on 16kHz, mono-channel audio. This acoustic model and language model pair will only be able to recognize the words "yes" and "no", which is a common use case in IVR systems.
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
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. The acoustic model comes in two forms: quantized and unquantized. There is a size<->accuracy trade-off for acoustic model quantization. For this combination of acoustic model and language model, we optimize for small size.
| Model type | Vocabulary | Filename | Size |
|---|---|---|---|
| Acoustic model | open | model_quantized.tflite |
46M |
| Language model | small | yesno.scorer |
640B |
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. Testing accuracy is reported for MLS and Librispeech.
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-large-vocab
English STT v1.0.0 (Large 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
STTissues, open a new discussion onSTTdiscussions, 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 language model with settings lm_alpha=0.49506138236732433 and lm_beta=0.11939819449850608 (found via lm_optimizer.py with Common Voice):
- Librispeech clean: WER: 5.2%, CER: 1.9%
- Librispeech other: WER: 15.0%, CER: 7.3%
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 | large_vocabulary.scorer |
127M |
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.
English STT v1.0.0-digits
English STT v1.0.0 (digits)
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:
Small vocabulary Speech-to-Text - Model version:
v1.0.0-digits - 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
STTissues, open a new discussion onSTTdiscussions, or chat with us on Gitter.
Intended use
Closed vocabulary (digits "zero" through "nine") Speech-to-Text for the English Language on 16kHz, mono-channel audio. This acoustic model and language model pair will only be able to recognize the words {"zero","one","two","three","four","five","six","seven","eight" and "nine"}, which is a common use case in IVR systems.
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
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. The acoustic model comes in two forms: quantized and unquantized. There is a size<->accuracy trade-off for acoustic model quantization. For this combination of acoustic model and language model, we optimize for small size.
| Model type | Vocabulary | Filename | Size |
|---|---|---|---|
| Acoustic model | open | model_quantized.tflite |
46M |
| Language model | small | digits.scorer |
1.7K |
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. Testing accuracy is reported for MLS and Librispeech.
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.
Swahili (Congo) STT v0.3.0
Swahili (Congo) STT v0.3.0 (Alp Öktem)
Jump to section:
- Model details
- Intended use
- Performance Factors
- Metrics
- Training data
- Training parameters
- Language models
- Evaluation data
- Ethical considerations
- Caveats and recommendations
Model details
- Person and organization developing model: Alp Öktem @Clear Global/Translators without Borders.
- Model language: Swahili (Congo) /
swc/sw-cd - Model date: August 26, 2021
- Model type:
Speech-to-Text - Model version:
v0.3.0 - Compatible with 🐸 STT version:
v0.10.0a13 - License: Custom (
LICENSE.txt) - Citation details:
@techreport{swc-stt, author = {\"Oktem, Alp}, title = {SWC STT 0.3}, institution = {Translators without Borders}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {June}, number = {STT-SWC-0.3} } - Official page: https://gamayun.translatorswb.org/data/
- Where to send questions or comments about the model: Directly to Alp Öktem or you can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text for the Congolese dialect of 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
The following Word Error Rates and Character Error Rates are reported on Congolese Swahili Commands dataset.
| Test Corpus | Scorer | WER | CER |
|---|---|---|---|
| TICO-19 devset | swc-general | 18.31% | 6.15% |
| Congolese Swahili Commands | swc-commands | 21.08% | 20.82% |
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
swc-stt-0.3.pbmm: 188.9 Mb
swc-stt-0.3.tflite:47.3 Mb
swc-general.scorer: 158.6 Mb
swc-commands.scorer: 2.9 Kb
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
Acoustic model was trained on top of English STT model using portions of Congolese Swahili audio mini-kit and TICO-19 Congolese Swahili testing set. It was converted to 16kHz WAV before training.
Total train size: 8.93 (mini-kit) + 3.27 (TICO-19 testset) = 12.2 hours
Dev size: 0.49 hours (mini-kit)
Test size: 1.71 hours (TICO-19 devset)
Training parameters
| Parameter | Value |
|---|---|
| Epochs | 200 |
| Drop source layers | 2 |
| Learning rate | 0.001 |
| Dropout rate | 0.2 |
| augment frequency_mask | [p=0.8,n=2:4,size=2:4] |
| augment time_mask | [p=0.8,n=2:4,size=10:50,domain=spectrogram] |
| Train/test/dev batch size | 32 |
Language models
Model is packaged with two language models (scorers):
- General purpose language model (
swc-general.scorer) is trained on a 37.7M word mixed Swahili text corpus - Commands language model (
swc-commands.scorer) is trained on 12 commands (numbers from 1 to 10 and yes/no) which are listed invocab-commands.txt.
Evaluation data
The Model was evaluated on two different sets:
- Congolese Swahili audio commands corpus: 185 sample subset (1.8 minutes) consisting of 5 speakers uttering numbers 1 to 10 and yes/no in Congolese Swahili. For this evaluation, the
swc-commandslanguage model was used. - Congolese Swahili TICO-19 audio development set: 536 sample subset (1.71 hours) consisting of TICO-19 domain sentences spoken by a male and female speaker (listed in
swc-tico-test.csv). For this evaluation, theswc-generallanguage model was used.
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 yesno-v0.0.1
English STT yesno-v0.0.1 (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: English / English /
en - Model date: July 26, 2021
- Model type:
Speech-to-Text/constrained vocabulary/yesno - Model version:
v0.0.1 - Compatible with 🐸 STT version:
v0.9.3 - License: Apache 2.0
- Citation details:
@techreport{english-yesno-stt, author = {Coqui}, title = {English yesno STT v0.0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {July}, number = {STT-EN-YESNO-0.0.1} } - Where to send questions or comments about the model: You can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text yesno model for the English Language on 16kHz, mono-channel audio. This model has been trained to only recognize the two words "yes" and "no" in English.
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 model was trained and evaluted on the Common Voice Target Segments Corpus, specifically, only on "yes" and "no" audio clips.
| Test Corpus | Word Error Rate |
|---|---|
| Common Voice 6.1 (Target Segments Corpus "yes" and "no") | 1.6% |
Model Size
yesno.pbmm: 319K
yesno.scorer: 1.7K
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
The model was trained and evaluted on the Common Voice Target Segments Corpus, specifically, only on "yes" and "no" audio clips.
Evaluation data
The model was trained and evaluted on the Common Voice Target Segments Corpus, specifically, only on "yes" and "no" audio clips.
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 v0.2.0
Czech STT v0.2.0 (Vojtěch Drábek)
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 Vojtěch Drábek.
- Model language: Czech / čeština /
cs - Model date: July 21, 2021
- Model type:
Speech-to-Text - Model version:
v0.2.0 - Compatible with 🐸 STT version:
v0.9.3 - License: CC-BY-NC
- Citation details:
@techreport{czech-stt, author = {Drábek,Vojtěch}, title = {Czech STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {July}, number = {STT-CS-0.2} } - Where to send questions or comments about the model: You can leave an issue on the model release page or
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
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 |
|---|---|---|
| Common Voice | 42.3% | 11.2% |
| Vystadial 2016 | 50.8% | 19.6% |
| Parliament Plenary Hearings | 21.5% | 5.2% |
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 the following corpora:
- Vystadial 2016 – Czech data
- OVM – Otázky Václava Moravce
- Czech Parliament Meetings
- Large Corpus of Czech Parliament Plenary Hearings
- Common Voice Czech
- Some private recordings and parts of audioboooks
Evaluation data
The model was evaluated on Common Voice Czech, Large Corpus of Czech Parliament Plenary Hearings and Vystadial 2016 – Czech data 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.
Bengali STT v0.1.0
Bengali STT v0.1.0 (Alp Öktem)
Jump to section:
- Model details
- Intended use
- Performance Factors
- Metrics
- Training data
- Training parameters
- Evaluation data
- Ethical considerations
- Caveats and recommendations
Model details
- Person and organization developing model: Alp Öktem @ Clear Global/Translators without Borders.
- Model language: Bengali / বাংলা /
bn/ben - Model date: June 9, 2021
- Model type:
Speech-to-Text - Model version:
v0.1.0 - Compatible with 🐸 STT version:
v0.10.0a6 - License: Custom
- Citation details:
@techreport{bengali-stt, author = {\"Oktem, Alp}, title = {Bengali STT 0.1}, institution = {Translators without Borders}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {June}, number = {STT-BN-0.1} } - Official page: https://gamayun.translatorswb.org/data/
- Where to send questions or comments about the model: You can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text for the Bengali 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 Large Bengali ASR training data set.
| Test Corpus | WER | CER |
|---|---|---|
| Large Bengali ASR training data set | 30.6% | 11.0% |
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: 189.3M
general-bn.scorer: 71.9M
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
Acoustic model was trained on top of English STT model using the Large Bengali ASR training data set. It was converted to 16kHz WAV before training.
Train size: 203,067 samples, 199.99 hours
Dev size: 10,690 samples, 10.55 hours
Language model was trained on OSCAR and Bengali portions of English-Bengali parallel corpora available from OPUS.
Lines: 782,827
Tokens: 13,953,256
Training parameters
| Parameter | Value |
|---|---|
| Epochs | 200 |
| Drop source layers | 2 |
| Learning rate | 0.001 |
| Dropout rate | 0.2 |
| augment frequency_mask | [p=0.8,n=2:4,size=2:4] |
| augment time_mask | [p=0.8,n=2:4,size=10:50,domain=spectrogram] |
| Train/test/dev batch size | 32 |
Evaluation data
The Model was evaluated on a 2000 sample subset (1.84 hours) of Large Bengali ASR training data set. Testing set filenames and transcriptions are included with the model.
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.
Russian STT v0.1.0
Russian STT v0.1.0 (Joe Meyer)
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: Russian / русский язык /
ru - Model date: May 12, 2021
- Model type:
Speech-to-Text - Model version:
v0.1.0 - Compatible with 🐸 STT version:
v0.9.3 - License: CC-0
- Citation details:
@techreport{russian-stt, author = {Meyer,Joe}, title = {Russian STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {May}, number = {STT-CV6.1-RU-0.1} } - Where to send questions or comments about the model: You can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text for the Russian 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 non-official held-out test set from Common Voice 6.1 with the use of an external language model. The official validated.tsv was re-processed by CorporaCreator to include all repeat sentences.
| Test Corpus | WER | CER |
|---|---|---|
| Common Voice | 32.3% | 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.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 a non-official training set from Common Voice 6.1. The official validated.tsv was re-processed by CorporaCreator to include all repeat sentences.
Evaluation data
This model was evaluated on a non-official testing set from Common Voice 6.1. The official validated.tsv was re-processed by CorporaCreator to include all repeat sentences.
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.
Dutch STT v0.0.1
Dutch STT v0.0.1 (acabunoc)
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 Abigail Cabunoc Mayes.
- Model language: Dutch / Nederlands /
nl - Model date: July 12, 2020
- Model type:
Speech-to-Text - Model version:
v0.0.1 - Compatible with 🐸 STT version:
v0.9.3 - License: MPL
- Citation details:
@techreport{dutch-stt, author = {Cabunoc Mayes,Abigail}, title = {Dutch STT 0.0.1}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2020}, month = {July}, number = {STT-CV-NL-0.0.1} } - Where to send questions or comments about the model: You can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text for the Dutch 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 using a language model: Github.
| Test Corpus | WER | CER |
|---|---|---|
| Common Voice 5.1 | 87.8% | 65.3% |
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: 660K
model.tflite: 221K
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 5.1 train.
Evaluation data
The Model was evaluated on Common Voice 5.1 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.
Yoruba STT v0.1.0
Yoruba STT v0.1.0 (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 Francis Tyers and the Inclusive Technology for Marginalised Languages group.
- Model language: Yoruba / Èdè Yorùbá /
yo - 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{yoruba-stt, author = {Tyers,Francis}, title = {Yoruba STT 0.1}, institution = {Coqui}, address = {\url{https://github.com/coqui-ai/STT-models}} year = {2021}, month = {April}, number = {STT-ALFFA-YO-0.1} } - Where to send questions or comments about the model: You can leave an issue on
STT-modelissues, open a new discussion onSTT-modeldiscussions, or chat with us on Gitter.
Intended use
Speech-to-Text for the Yoruba 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 | 71.6% | 23.0% |
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 the Yoruba subset of the ALFFA corpus.
Evaluation data
The Model was evaluated on the Yoruba 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.