Tools for automated transcription of audio and video fragments
Speech recognition, speech-to-text (STT) or automated speech recognition (ASR) is a technology that makes it possible to convert spoken text in videos or audio into text, such as the automatic subtitles on YouTube or Zoom. In this overview we focus on transcribing interviews for oral history, but virtual assistants like Siri or Google Assistant are also a form of this technology.
Speech recognition is a relatively old technology. The first commercial tools appeared in the early 1990s. They use models, systems that are trained on a certain set of data to recognize patterns and make decisions without human intervention. Speech recognition models are language models trained on audio such as interviews, audiobooks, lectures and presentations. The strength of the speech recognition tool depends enormously on the model used.
Possibilities of speech recognition tools
A research by meemoo in 2020-2021 showed that speech recognition technology was less good at transcribing audio from conversations, but great strides have been made in recent years which has greatly improved the technology, partly due to the greater computing power of computers, the progress in technology related to machine learning and big data and the improved language models. This has ensured that the tools generate more natural language and produce fewer nonsensical texts, thus improving transcriptions of conversations enormously.
In addition, speech recognition tools can do more than just transcribe text. They can also:
- recognize different speakers and indicate which text was spoken by which speaker
- indicate and remove filler words such as euhm from the text
- indicate silences
- create summaries and translations
- ...
Several tools also have a possibility to add a dictionary. Here you can set specialized words that would otherwise be transcribed incorrectly.
Points of attention when using the tools
You can take into account a number of factors to achieve better results:
- The strength of the speech recognition tools depends on the model used: e.g., which languages does the model support and how well can the model handle dialects or colloquial language? The choice for the model depends on the needs and use cases, such as the trade-off between speed and accuracy, and the language that must be transcribed. In the overview we chose accuracy (it is important that the text is correct than that the tool is fast) and Dutch as colloquial language. They were also all tested on interviews in which slight dialect or intermediate language was spoken. The most accurate tools scored well for this, but generally it was noted that interviews with pronounced dialects resulted in poorer transcriptions.
- Another important point of attention is the quality of the recording. Recordings with clear sound and without background noise give better results than recordings with poor sound quality (e.g. noise) and background noise. Speakers who speak clearly are also transcribed better than speakers who mumble.
- Desktop applications use the computing power of the computer when transcribing the text. If your computer has a dedicated GPU, the transcription will proceed much faster. A dedicated GPU is a special GPU with its own card connected to the motherboard, while an integrated GPU is embedded in the same chip as the CPU. This article (in Dutch) explains more about the different GPUs. If your computer does not have this GPU, then the CPU or processor is used. This is slower. You can also use an online service so that you are not limited by the limits of your computer.
Overview of the different tools
At the request of heritage organisations for support in writing out interviews, a number of tools were tested. Most tools can do more than just transcribe, but those functions were not investigated deeper.
Amberscript
Amberscript is a commercial web platform that allows you to create transcriptions and subtitles for audio and video. When you want to transcribe audio or video, you must upload those fragments into their web environment. Which model is used is not documented.
It supports:
- different languages, including Dutch;
- speaker identification;
- and indication of timestamps
Advantages:
- Editing transcriptions is very clear.
- Possibility to add a dictionary.
Disadvantages:
- Speaker identification is inaccurate when there are pauses in the audio fragment.
- Pauses are not indicated.
- Accuracy in Dutch is not good. The transcribed text contains quite a few errors.
Using Amberscript costs money. If you want to pay once, it will cost you €15/hour, whereby you can purchase a maximum of hunderd hours at a time. It is also possible to take out a subscription. You then pay €40/month per five hours of uploaded audio or video if you take a yearly subscription, or €50/month for five hours of audio or video. It is not possible to transfer unused hours to another month.
Audapolis
Audapolis is an open source and offline desktop application that uses the Vosk model. Vosk is a relatively small, but also older model that is mainly used for chatbots, smart home applications and virtual assistants. It was originally developed for smaller devices such as smartphones and microcomputers.
The tool can transcribe twenty languages, including Dutch, identifies which text was spoken by which speaker and also indicates timestamps. Because the tool works offline, you can also use the tool, for example, on the train when traveling home after an interview.
Advantages:
- free and open source application;
- works offline;
- includes a built-in editor to improve the transcribed text;
- detects and removes euhms and pauses.
Disadvantages:
- does not work well with accents or (slight) dialects;
- accuracy of transcribed text and speaker identification is low;
- text can only be exported in HTML format;
- tool hasn't been updated for a year.
Limecraft
Limecraft is an online platform for managing audio and video and offers the possibility to transcribe that content as an extra feature. It uses six models for this: Vocapia, Speechmatics, Google Speech, Microsoft Auze, Scriptix and Kaldi. As a user you cannot choose which of those models you want to use. The platform makes that choice for you based on the purposes and the language you choose. Because so many models are used, it can transcribe more than hunderd languages, including Dutch. Just like the previous tools, it can identify speakers and indicate timestamps. On the platform, multiple people can collaborate on a transcription, like working together on a document in Sharepoint or Google Drive.
Advantages:
- user-friendly interface with simple editing possibilities;
- fast;
- extensive export options (.pdf, .csv, .doc);
- possibility to add your own dictionary;
- has extra features, such as creating subtitles, topic detection and making summaries;
- it is a Belgian company, which means support and contact are in Dutch and the same time zone.
Disadvantages:
- Uhms are not recognized well and silences are not indicated.
- The distinction between speakers is not always good, but you can edit this.
- It makes strange things of words that it does not know.
- Transcription is an extra feature, making the platform quite expensive and having many functionalities you do not need if you only want to transcribe.
The prices range from free (one user with 5h of material), €85/month (five users with 25h of material) or €275/month (for larger teams). To have audio and video transcribed, you pay an extra €15/hour. Limecraft also offers the possibility to translate that transcription. That also costs €15/hour.
Sonix
Sonix is also a commercial web platform where you can collaborate on transcriptions. There are extensive possibilities to edit transcriptions, adjust timestamps, etc. It can transcribe more than 49 languages, including Dutch, recognizes the different speakers well and indicates timestamps. Finally, it also has the ability to indicate in color codes how certain Sonix is about certain transcriptions.
Advantages:
- user-friendly interface with extensive and simple editing possibilities;
- transcribes quickly;
- possibility to add your own dictionary for specific words;
- extensive export options
- has an extra (paid) feature to create summaries of transcriptions
Disadvantages:
- Uhms are not recognized well.
- It compresses the original media files when using the cheapest tariff plan, so you can no longer export the original media files.
- With the cheapest tariff plan you only have support per email.
Sonix has different subscription models:
- Standard pay-as-you-go where you pay \$10/hour audio or video
- Premium for organizations that regularly want to transcribe audio and video and need more collaboration possibilities. For this you pay \$5/hour audio or video and monthly \$22 per user.
- Enterprise for high volumes of transcription needs with extensive collaboration options and content analysis.
Je kan de mogelijkheden van Sonix gratis testen op 30 minuten aan audio en video.
Speechmatics
Speechmatics is een bedrijf dat zelf een eigen gesloten spraakherkenningsmodel ontwikkeld heeft en tegen betaling API's en een platform aanbiedt om audio en video te laten transcriberen en vertalen. Ze proberen te concurreren met grote bedrijven zoals Google, Amazon en Microsoft en volgens testen scoort hun model beter dan die techbedrijven. Het is te gebruiken voor zowel opgenomen media als voor real-time audio en video. De software kan 52 talen transcriberen, waaronder Nederlands, sprekers identificeren en duidt tijdscodes aan. Speechmatics focust zich enkel op transcriptie. Mediabestanden en hun transcripties worden daardoor maar voor een week bewaard op het platform. Dat heeft als voordeel dat ze een van de goedkoopste speech-to-text aanbieders zijn.
Voordelen:
- zeer accuraat;
- verwijdert euhms;
- platform focust zich enkel op transcriptie, waardoor je niet voor overbodige bells and whistles moet betalen;
- exporteert naar platte tekst (.txt), SRT (voor ondertitels) en JSON.
Nadelen:
- hallucineert op termen die het niet kent;
- duidt stiltes niet aan;
- geen tijdsaanduidingen wanneer je exporteert naar platte tekst of wanneer je de kopieerfunctie gebruikt;
- mediabestanden en transcripties worden maar een week bewaard op het webplatform;
- het webplatform heeft moeite met het opladen van video.
Speechmatics heeft geen abonnementsformule. Je kan maandelijks gratis 4u opgeladen audio of video laten transcriberen (en ook nog 4u real-time audio en video). Daarnaast betaal je per uur en is de prijs afhankelijk van de gewenste accuraatheid van de getranscribeerde tekst. Je betaalt $0,8/uur voor de standard accuracy en $1,04/uur voor de enhanced accuracy of het meest accurate model.
noScribe
NoScribe is een gratis opensourcetool voor het transcriberen van audio en video. Het is een offline desktopapplicatie die voor transcriptie gebruikt maakt van het Whisper-model (voor meer info over Whisper, zie onder) van OpenAI, het bedrijf dat ook ChatGTP ontwikkeld heeft. NoScribe kan meer dan 99 talen transcriberen, waaronder Nederlands, identificeert sprekers en geeft tijdscodes aan. Het maakt (nog) geen gebruik van het meest recente (en voor het Nederlands meest accurate) model van Whisper, omdat die versie van dat model minder goed scoort op sommige andere talen.
Voordelen:
- gratis en opensource;
- zeer accuraat, hallucineert minder en blijft consistent bij termen die het niet kent;
- beschikt over editorsoftware om transcripties te verbeteren;
- kan exporteren naar HTML, platte tekst (.txt) en VTT (voor ondertitels);
- kan een dedicated GPU gebruiken om de transcriptie sneller te laten verlopen.
Nadelen:
- snelheid van transcriptie is afhankelijk van de eigen computer;
- kan hallucineren op stilte, al konden we dat in de praktijk niet vaststellen;
- meertalige audio (bv. een interview waarin verschillende talen gesproken worden) is niet ondersteund;
- maakt wel eens fouten bij het herkennen van sprekers.
Whisper
Whisper is een model voor spraakherkenning, ontwikkeld door OpenAI, dat voor het eerst als opensourcesoftware uitgebracht werd in 2022. Het kan gebruikt worden om verschillende talen te transcriberen en om verschillende talen te laten vertalen naar Engels. Het is ingebouwd in verschillende spraakherkenningstools, zoals noScribe, maar kan ook gebruikt worden als commandlinetool. Whisper kan honderd talen transcriberen, waaronder Nederlands, en duidt tijdscodes aan. Uit testen op twee datasets blijkt ook dat de laatste versie van Whisper zeer goed scoort op Nederlands.
Voordelen:
- opensource en gratis
- zeer accuraat; de commandlinetool gebruikt het meest recentste model, wat voor het Nederlands ook het meest accurate model is
- exporst zijn mogelijk in platte tekst (.txt), SRT (ondertiteling), VTT (ondertiteling), TSV (een tabulair formaat vergelijkbaar met CSV) en JSON
Nadelen:
- snelheid is afhankelijk van de eigen computer, zeker als je geen dedicated GPU hebt (zie noScribe) kan de transcriptie zeer langzaam gaan (maar nog steeds sneller dan als je het zelf zou doen)
- duidt geen stiltes aan
- Whisper kan hallucineren op stilte, maar het is mogelijk om via de commandline hier aanpassingen aan te doen
- meertalige audio is niet ondersteund
- geen intuïtieve grafische interface (GUI); enkel te gebruiken via de command line.
- geen omgeving waarin je de transcriptie kan verbeteren.
Conclusie
Er zijn, afhankelijk van je noden, verschillende tools die je kan gebruiken voor het geautomatsieerd transcriberen van audio- en videofragmenten. Om eenvoudiger een keuze te maken, kan je onderstaande tabel gebruiken. In de tabel wordt onder meer aangeduid welke mogelijkheden de tools volledig (aangeduid met X) of gedeeltelijk (aangeduid met /) ondersteunen en hun prijscategorie: €0 betekent gratis, € staat voor een prijs minder dan €5/uur, €€ geeft een prijs tussen €5/uur en €15/uur weer en €€€ is een prijs hoger dan €15/uur.
| Amberscript | Audapolis | Limecraft | Sonix | Speechmatics | noScribe | Whisper | |
|---|---|---|---|---|---|---|---|
| Ondersteunt Nederlands | X | X | X | X | X | X | X |
| Accuraat | X | X | X | X | X | ||
| Gebruiksvriendelijk | X | X | X | X | X | X | |
| Snel | X | X | X | X | |||
| Mogelijkheid tot verbeteren van transcriptie | X | X | X | X | X | ||
| Mogelijkheid tot samenwerken aan transcriptie | X | X | |||||
| Identificeert sprekers | X | X | X | X | X | X | |
| Detecteert euhms | X | / | X | X | X | ||
| Duidt tijdscodes aan | X | X | X | / | X | X | |
| Detecteert pauzes | X | X | X | X | X | ||
| Mogelijkheid tot toevoegen eigen woorden | X | X | |||||
| Exportformaten | .csv, .doc, .json, .rtf, .srt, .stl, .txt, .vtt | .html | .docx, .pdf, .srt, .txt, .vtt | .csv, .doc, .pdf | .json, .srt, .txt | .html, .txt, .vtt | .csv, .json, .srt, .tsv, .txt, .vtt |
| Open source | X | X | X | ||||
| Cloudservice | X | X | X | X | |||
| Prijs | €€ | €0 | €€€ | €€ | € | €0 | €0 |
Auteur: Nastasia Vanderperren (Meemoo, Vlaams instituut voor het archief) en Lode Scheers (Meemoo, Vlaams instituut voor het archief)