14 new CrewChief voices + make your own

[cktl]

Hi, I'm happy to share these 14 full alternate voice packs, with over 30,000 .wav files each... alongside a free tool to automatically create a custom CrewChief voicepack based on your voice.

https://github.com/cktlco/crew-chief-autovoicepack


The voice packs contain imperfections when compared to the default "Jim" voice -- in particular, speed/pace/rhythm may seem "too slow" or emotionally misaligned, there are frequent mispronounciations, and a small percent of the generated audio clips have "corrupted" audio. Some of these things can be manually improved, with additional notes in the README.

Nonetheless, the results are more than good enough to use and share -- I've been using these as my daily driver voices for several months now and it's been a welcome and fun alternative despite the quality gap to the official "Jim" recordings.

Thanks!

[Seven Smiles]

Oh wow! Great work. There has been talk of doing this for a long time, good to see someone pull it off

[Sparten]

Thanks for posting it here, i saw you posted it on reddit a few days ago, I haven't had the time to dig into it yet but some of the samples sounds are not to bad.

[Berti77]

Hello, does anyone have a language pack in German? Thank you

[cktl]

Hello, does anyone have a language pack in German? Thank you

Hi Berti77, I put some notes in the README about how someone might translate the text phrases used to generate a voicepack. The speech generation model does a nice job pronouncing French, German, and other language corner names at least, so it would work after some manual translation effort.

https://github.com/cktlco/crew-chief...erent-language

English:

- Human or Machine-translate the `original_text` column of `phrase_inventory.csv` using an LLM like ChatGPT. This can be done iteratively, copy/pasting an acceptable number of rows at a time to the LLM, or fully automated via API calls.

- After translation, manually review the csv file in a tool like Excel or LibreOffice Calc

- Manually edit any incorrect translations

- Optionally, tweak the TTS pronunciation column `text_for_tts` for troublesome entries

- Generate the voice pack using the normal process

- Important: ensure that your updated `phrase_inventory.csv` file is mounted into the container using the `docker run ... -v ...` command line parameter.
- For example, use `docker run ... -v /path/to/MY_CUSTOM_phrase_inventory.csv:/app/phrase_inventory.csv ...` to override the default phrases with your translated phrases.

- Let the voice pack generate for a few minutes

- Review the output audio files to judge how well it worked

The coqui xtts model is multilingual and should work well with many languages, but the quality of the translations will depend on the quality of the machine translation and the quality of the original audio recordings. You may need to experiment with different translation techniques to get the best results, but crew-chief-autovoicepack will still do most of the work

---

German:

Maschinenübersetzen der original_text-Spalte aus der Datei phrase_inventory.csv mit einem LLM wie ChatGPT. Dies kann iterativ geschehen, indem man eine akzeptable Anzahl von Zeilen gleichzeitig in das LLM kopiert, oder vollständig automatisiert über API-Aufrufe.

Nach der Übersetzung die csv-Datei in einem Tool wie Excel oder LibreOffice Calc manuell überprüfen.

Eventuell falsche Übersetzungen manuell korrigieren.

Optional die Spalte text_for_tts zur TTS-Aussprache bei problematischen Einträgen anpassen.

Das Sprachpaket wie gewohnt erstellen.

Wichtig: Stellen Sie sicher, dass Ihre aktualisierte phrase_inventory.csv-Datei mit dem Befehl docker run ... -v ... in den Container eingebunden wird.

Verwenden Sie zum Beispiel docker run ... -v /path/to/MY_CUSTOM_phrase_inventory.csv:/app/phrase_inventory.csv ..., um die Standardphrasen durch Ihre übersetzten Phrasen zu ersetzen.


Lassen Sie das Sprachpaket einige Minuten generieren.

Überprüfen Sie die ausgegebenen Audiodateien, um zu beurteilen, wie gut es funktioniert hat.

Das coqui xtts Modell ist mehrsprachig und sollte mit vielen Sprachen gut funktionieren, aber die Qualität der Übersetzungen hängt von der Qualität der maschinellen Übersetzung und der ursprünglichen Audioaufnahmen ab. Möglicherweise müssen Sie mit verschiedenen Übersetzungstechniken experimentieren, um die besten Ergebnisse zu erzielen, aber crew-chief-autovoicepack erledigt dennoch den Großteil der Arbeit — stellen Sie sich vor, Sie müssten zehntausend Audioclips selbst aufnehmen und bearbeiten

[Berti77]

Hi Berti77, I put some notes in the README about how someone might translate the text phrases used to generate a voicepack. The speech generation model does a nice job pronouncing French, German, and other language corner names at least, so it would work after some manual translation effort.

https://github.com/cktlco/crew-chief...erent-language

English:

- Human or Machine-translate the `original_text` column of `phrase_inventory.csv` using an LLM like ChatGPT. This can be done iteratively, copy/pasting an acceptable number of rows at a time to the LLM, or fully automated via API calls.

- After translation, manually review the csv file in a tool like Excel or LibreOffice Calc

- Manually edit any incorrect translations

- Optionally, tweak the TTS pronunciation column `text_for_tts` for troublesome entries

- Generate the voice pack using the normal process

- Important: ensure that your updated `phrase_inventory.csv` file is mounted into the container using the `docker run ... -v ...` command line parameter.
- For example, use `docker run ... -v /path/to/MY_CUSTOM_phrase_inventory.csv:/app/phrase_inventory.csv ...` to override the default phrases with your translated phrases.

- Let the voice pack generate for a few minutes

- Review the output audio files to judge how well it worked

The coqui xtts model is multilingual and should work well with many languages, but the quality of the translations will depend on the quality of the machine translation and the quality of the original audio recordings. You may need to experiment with different translation techniques to get the best results, but crew-chief-autovoicepack will still do most of the work

---

German:

Maschinenübersetzen der original_text-Spalte aus der Datei phrase_inventory.csv mit einem LLM wie ChatGPT. Dies kann iterativ geschehen, indem man eine akzeptable Anzahl von Zeilen gleichzeitig in das LLM kopiert, oder vollständig automatisiert über API-Aufrufe.

Nach der Übersetzung die csv-Datei in einem Tool wie Excel oder LibreOffice Calc manuell überprüfen.

Eventuell falsche Übersetzungen manuell korrigieren.

Optional die Spalte text_for_tts zur TTS-Aussprache bei problematischen Einträgen anpassen.

Das Sprachpaket wie gewohnt erstellen.

Wichtig: Stellen Sie sicher, dass Ihre aktualisierte phrase_inventory.csv-Datei mit dem Befehl docker run ... -v ... in den Container eingebunden wird.

Verwenden Sie zum Beispiel docker run ... -v /path/to/MY_CUSTOM_phrase_inventory.csv:/app/phrase_inventory.csv ..., um die Standardphrasen durch Ihre übersetzten Phrasen zu ersetzen.


Lassen Sie das Sprachpaket einige Minuten generieren.

Überprüfen Sie die ausgegebenen Audiodateien, um zu beurteilen, wie gut es funktioniert hat.

Das coqui xtts Modell ist mehrsprachig und sollte mit vielen Sprachen gut funktionieren, aber die Qualität der Übersetzungen hängt von der Qualität der maschinellen Übersetzung und der ursprünglichen Audioaufnahmen ab. Möglicherweise müssen Sie mit verschiedenen Übersetzungstechniken experimentieren, um die besten Ergebnisse zu erzielen, aber crew-chief-autovoicepack erledigt dennoch den Großteil der Arbeit — stellen Sie sich vor, Sie müssten zehntausend Audioclips selbst aufnehmen und bearbeiten

Hello, thank you for your answer, but unfortunately I have no idea what you are talking about. Is there perhaps a YouTube video that explains this? Thanks

[cktl]

Hello, thank you for your answer, but unfortunately I have no idea what you are talking about. Is there perhaps a YouTube video that explains this? Thanks

Hi, no video but the process is documented pretty exhaustively in the linked README. Which part are you stuck on?

Admittedly, if you are not used to working with git repos or using docker or other software engineering tasks it will seem a bit daunting. If anyone wants to make a video, it would also be welcome.

[Berti77]

Hi, no video but the process is documented pretty exhaustively in the linked README. Which part are you stuck on?

Admittedly, if you are not used to working with git repos or using docker or other software engineering tasks it will seem a bit daunting. If anyone wants to make a video, it would also be welcome.

It fails when I download the image. I downloaded the Docker and now I don't know what to do.

what should I do with it?

docker pull ghcr.io/cktlco/crew-chief-autovoicepack:latest

I can't find a file in the Docker under Container or Image when I copy it in



I think it's not for me, I have no idea how it should work.
Is it possible for someone to create a German audio voice for Crewchief for me? I would also show my gratitude, thank you

[cktl]

Hi @Berti77, great job for trying to get the process working. It will only be easy for someone who is very familiar with the details, so I'm glad you tried.

Good news! To help, I've added multi-language support to crew-chief-autovoicepack and I added an experimental German language voicepack.

Let's discuss in this thread: https://thecrewchief.org/showthread....autovoicepack)

[cktl]

Hi all, I've released an updated version of crew-chief-autovoicepack. The 13 English voice packs were regenerated to adopt the improvements below, alongside 3 new German language voices. To update, just [install] over any previous versions.

Download: https://github.com/cktlco/crew-chief...eases/tag/v3.0


- Greatly Improved Voice Quality
Previous voicepacks suffer from occasional "haunted" or "buzzing" audio artifacts which are annoying and immersion-breaking. In this release, the unwanted non-speech audio is automatically detected and regenerated again until valid. This won't fix genuine mispronunciations, wrong accents, or similar natural speech issues.

This was accomplished via creating a new ML model to classify the speech output as good or bad: [xtts-integrity]

- Faster-paced speech
A top community request is to increase the "speaking rate" from the current slightly-too-slow pace, including reducing any unnatural pauses between words.

Due to the additional guardrails added by the xtts-integrity integration, I've been able to experiment with more extreme values for the xtts model's speed parameter. Previously, I used a speed of 1.2 as it was a good balance between natural-sounding output and "stability" of the output -- for example, higher speed values would generate far more of the buzzing audio artifacts referred to above. Now that corrupt audio clips are auto-filtered out, we are free to use any other values. I chose 1.6 as the new default.

Experiment with other values using the --xtts_speed command-line parameter to generate_voice_pack.py. For example, 1.6 and 1.9 work fine and are noticeably faster-paced than each other, with 1.9 being faster than a normal human speaker.

- Smaller file sizes
To save a bit of space, the .wav files being produced now follow the same format as the official Crew Chief "Jim" voice. This means 22KHz sample rate instead of 24KHz, and bit rate of 352 kb/s instead of 384 kb/s. The new faster speaking rate also means shorter files and thus notably lower file sizes.