Speech to text has been improved a lot in 2023 all thanks to research in AI. In this blog post, I will share the best speech-to-text API that you can use for your next project.
We will only provide API not open source model. The biggest advantage of using API over open source is
- It is easy to setup
- No infrastructure cost initially
- No problem in Scalability
Here are the best Speech-to-text APIs.
1. OpenAI Whisper
OpenAI recently introduced Whisper V3, a cutting-edge Speech to Text API that’s making waves in the tech world. What’s exciting about Whisper V3? It’s designed to understand and transcribe speech with incredible accuracy. This means it can turn spoken words into written text much better than before.
One important thing to note about Whisper V3 is its file size limit. Currently, it can handle audio files up to 25MB. This is something to keep in mind, especially if you’re dealing with longer recordings. It’s a crucial detail for anyone planning to use this API for large audio files.
Now, let’s talk about pricing. When it comes to cost, Whisper V3 offers competitive pricing options. However, OpenAI hasn’t publicly disclosed the exact pricing details yet. Whisper v2 currently cost $0.006/minute.
2. DataCrunch
OpenAI launched Whisper as open source model. That’s why there are so many whisper API providers. Datacrunch offers the Whisper-large-v2 API, a powerful tool for transforming spoken words into text.
Data crunch API is easy to use and affordable as well, probably the cheapest Whisper API. It will only cost you $0.0010/minute.
API implementation is also pretty easy. You can implement using curl request. Datacrunch API is also faster than OpenAI.
The file size limit of datacrunch is 200 MB which is much larger than the OpenAI file size (25 MB)
Datacrunch Whisper only accepts audio files.
3. Deepgram
Deepgram’s Nova 2 model is praised for its speed and accuracy in speech recognition. It’s trained on a vast amount of data, which helps in delivering better accuracy compared to many other available solutions.
The service supports speech recognition in over 30 languages, making it a versatile choice for global applications. This extensive language support is a significant advantage for users dealing with multi-lingual contexts.
Deepgram’s API includes features such as speaker diarization, smart formatting, automatic language detection, deep search, keyword boosting, multichannel support, and redaction of sensitive information like PCI and PII. These features enhance the overall utility and security of the API.
Deepgram offers competitive pricing for its Nova-2 model. The cost is $0.0043/min for pre-recorded content and streaming, which is more affordable than OpenAI’s Whisper API at $0.006/min. For higher usage (minimum 4k/year), the pricing remains the same, making it a cost-effective choice for users with substantial requirements.
Deepgram’s commitment to offering fast, accurate, and affordable speech recognition services makes its Nova 2 model an excellent choice for various applications, from simple transcriptions to complex multilingual projects.
For more detailed information about the model, language support, and specific features, you can refer to Deepgram’s developer documentation here.
4. Assembly AI
Assembly AI has made significant strides in the field of speech recognition with its Conformer 2 model, building upon its predecessor, Conformer 1. Here’s an overview of its features and capabilities:
Assembly AI’s Conformer 2: Advanced Speech Recognition Model
Enhanced Training and Performance
- Extensive Training Data: Conformer 2 is trained on 1.1 million hours of English audio data. This massive dataset has enabled significant improvements in recognizing proper nouns, alphanumerics, and noise robustness.
- Improvements Over Conformer 1: While maintaining similar word error rates as Conformer 1, Conformer 2 has achieved a 31.7% improvement in alphanumeric recognition, a 6.8% improvement in Proper Noun Error Rate, and a 12.0% improvement in noise robustness. These enhancements are crucial for real-world applications where accuracy in these areas is vital.
Speed and Efficiency
- Reduced Processing Time: One of the standout improvements in Conformer 2 is its speed. The model has significantly reduced the processing duration across various audio file lengths. For instance, transcription time for an hour-long file has been reduced from 4.01 minutes to 1.85 minutes, marking a substantial improvement in efficiency and making it one of the fastest options available.
Application in Generative AI
- Generative AI Applications: Conformer 2’s accuracy and speed make it an ideal choice for developing generative AI applications that leverage spoken data. The model’s enhanced capabilities enable product and development teams to incorporate highly accurate speech-to-text components in their AI pipelines.
In summary, Assembly AI’s Conformer 2 model represents a significant advancement in automatic speech recognition technology. With its extensive training on a large dataset, improvements in key performance metrics, and reduced processing time, it stands out as a top choice for various applications, especially those requiring high accuracy and speed in speech-to-text conversion.
5. Google Chrip
Google’s latest entry into the speech recognition market, the Universal Speech Model (USM), also referred to as Google Chirp, represents a significant advancement in the field. Here are the key details about this new model
USM is a family of state-of-the-art speech models with 2 billion parameters, trained on 12 million hours of speech and 28 billion sentences of text. This extensive training enables the model to recognize a wide range of languages and dialects, including under-represented ones.
Support for Over 300 Languages: The USM is designed to handle automatic speech recognition (ASR) across more than 300 languages. This includes not only widely spoken languages like English and Mandarin but also less common languages like Amharic, Cebuano, Assamese, and Azerbaijani.
Enhanced Accuracy: The USM has shown to have a lower word error rate (WER) compared to other models. For example, in English (en-US), USM achieves a 6% relative lower WER compared to Google’s internal state-of-the-art model. When compared to Whisper’s large-v2 model, which was trained with over 400k hours of labeled data, USM exhibits, on average, a 32.7% relative lower WER for the 18 languages where Whisper has less than 40% WER.
Google Chirp aims to offer more accurate and affordable speech recognition technology than previous models, although it’s still in beta and not available for all languages. The use of customer data to improve the model’s performance is a notable aspect of this technology, with the option for users to opt-out of data usage for training at an additional cost. This model represents a significant step forward in making speech recognition technology more inclusive and accessible to a broader range of languages and dialects.
6. Speechmatics
Speechmatics is renowned for its top-level speech-to-text API services. It provides both live transcription and translation across an impressive array of 49 languages. This wide language support positions Speechmatics as a versatile tool for global users.
A key feature of Speechmatics is its free plan, which offers up to 8 hours of transcription each month, divided equally between batch and real-time transcription. This plan is particularly beneficial for users with small workloads or businesses exploring large-volume usage.
In addition to basic transcription, Speechmatics extends its capabilities to include summarization, sentiment analysis, and topic detection. These advanced features enhance the utility of the API, making it suitable for a range of applications beyond simple transcription.
Regarding pricing, Speechmatics offers:
- Batch Transcription: Starting at $0.30/hr for Lite Mode standard accuracy, with higher rates for enhanced accuracy.
- Real-Time Transcription: Starting at $1.04/hr for standard accuracy, with a slightly higher rate for enhanced accuracy.
- Additional Services: Charges for translation are an additional $0.65/hr, while summarization, sentiment analysis, and topic detection are available at lower additional costs.
7. AWS Transcribe
AWS Transcribe, Amazon’s automatic speech recognition service, offers various features that make it a versatile tool for converting speech to text. While it may not be as accurate as some other models, it still provides several useful functionalities:
Key Features of AWS Transcribe
- Versatile Input Options: AWS Transcribe processes both live and recorded audio or video inputs, delivering high-quality transcriptions for search and analysis. This flexibility makes it suitable for a wide range of applications, from media content to call centres.
- Real-Time and Pre-Recorded Transcription: The service supports both streaming for real-time transcription and batch transcription for existing audio recordings. This dual capability allows for a variety of use cases, such as live events or processing archived recordings
- Adaptation to Various Audio Types: AWS Transcribe offers models specifically tuned to different audio types, like telephone calls or multimedia content. This specialization ensures better transcription accuracy in domain-specific scenarios
- Automatic Language Identification: The service can automatically identify the dominant language in an audio file, which is beneficial when dealing with multilingual content. This feature aids in accurate transcription across different languages
- Punctuation and Number Normalization: Transcribe automatically adds punctuation and formats numbers, making the transcripts easy to read and review. This feature enhances the quality of the output, making it comparable to manual transcription
Additional Features
- Timestamp Generation and Speaker Recognition: The service generates timestamps for each word and recognizes speaker changes, which is crucial for accurately capturing dialogues in scenarios like meetings or TV shows【92†source】.
- Channel Identification: Particularly useful for contact centers, AWS Transcribe can process a single audio file and automatically produce a transcript with channel labels【93†source】.
- Customization Options: Users can customize transcripts with their specific business needs and vernacular, including adding custom vocabulary for domain-specific words and phrases【94†source】.
- Custom Language Models: For enhanced speech recognition accuracy, AWS Transcribe allows the creation of custom language models tailored to specific use cases and domains【95†source】.
- Privacy and Safety Features: The service includes options to mask or remove sensitive words from transcription results, ensuring customer privacy and content appropriateness【96†source】.
While AWS Transcribe may not match the accuracy of some newer models like Google’s USM or OpenAI’s Whisper, its range of features, including domain-specific models, language support, and customization options, make it a valuable tool for various speech-to-text applications.