Audio Analysis Scripts

Tools for converting lecture recordings to text, identifying speakers, translating Czech to English, and redacting personal information — all runnable locally.

Typical workflow

Záznam.m4a
    │
    ▼  m4a_to_text.sh          (ffmpeg: M4A → MP3)
    │
    ▼  neural_sound_to_text.py  OR  neural_sound_to_text_2_entities.py
    │                              (Whisper transcription, optional speaker labels)
    │
    ▼  text_full.txt
    │
    ▼  simple_local_translator.py   (Ollama: Czech → English)
    │
    ▼  simple_local_anonymization.py   (optional: redact PII)

For a single-speaker recording, use neural_sound_to_text.py. For interviews or multi-speaker audio, use neural_sound_to_text_2_entities.py.

---

Prerequisites

System tools

Tool	Used by
ffmpeg	m4a_to_text.sh, audio_renamer_multiple.py
Ollama	simple_local_translator.py, ollm_*.sh

Python packages

Install into a virtual environment (Python 3.10+ recommended):

pip install torch transformers librosa soundfile tqdm python-dotenv ollama eyed3 numpy

For speaker diarization (neural_sound_to_text_2_entities.py), also install:

pip install pyannote.audio

On first run, Whisper (openai/whisper-large-v3) and Piiranha (iiiorg/piiranha-v1-detect-personal-information) models are downloaded from Hugging Face automatically.

---

Environment variables

Copy the example file and fill in your tokens:

cp .env.example .env

Variable	Required by	Description
HF_TOKEN	neural_sound_to_text_2_entities.py	Hugging Face access token. Accept the model terms for pyannote/speaker-diarization-3.1 before use.
MISTRAL_API_KEY	—	Reserved for future use; not currently read by any script.

---

Scripts

Audio conversion

m4a_to_text.sh

Converts Záznam.m4a to Záznam.mp3 using ffmpeg. Edit the filename inside the script if your input file has a different name.

bash m4a_to_text.sh

---

Transcription (Whisper)

Both scripts use openai/whisper-large-v3 and write output to text_full.txt (appended, not overwritten).

neural_sound_to_text.py

Chunked transcription for long audio files. Splits audio into 30-second segments with overlap to avoid cut-off words. No speaker identification.

python neural_sound_to_text.py --audio_file Záznam.mp3
python neural_sound_to_text.py --audio_file Záznam.mp3 --english
python neural_sound_to_text.py --audio_file Záznam.mp3 --volume_boost 1.5

Flag	Default	Description
--audio_file	Záznam.mp3	Path to the audio file
--english	off	Transcribe in English instead of Czech
--volume_boost	1.3	Volume multiplier after normalization

Uses Apple Silicon MPS, NVIDIA CUDA, or CPU automatically.

neural_sound_to_text_2_entities.py

Transcription with speaker diarization via pyannote.audio. Output labels each segment with a speaker ID and timestamp:

SPEAKER_00 [00:12 - 00:45]: ...
SPEAKER_01 [00:46 - 01:20]: ...

python neural_sound_to_text_2_entities.py --audio_file Záznam.mp3
python neural_sound_to_text_2_entities.py --audio_file Záznam.mp3 --num_speakers 3
python neural_sound_to_text_2_entities.py --audio_file Záznam.mp3 --output transcript.txt

Flag	Default	Description
--audio_file	Záznam.mp3	Path to the audio file
--english	off	Transcribe in English instead of Czech
--volume_boost	2.0	Volume multiplier after normalization
--hf_token	from .env	Hugging Face token (overrides HF_TOKEN)
--num_speakers	2	Expected number of speakers
--output	text_full.txt	Output file path

---

Post-processing

simple_local_translator.py

Translates text_full.txt from Czech to English using a local Ollama model (translategemma:12b by default). Pull the model first:

ollama pull translategemma:12b
python simple_local_translator.py

Edit input_file and model at the top of the script to change the source file or model. Translation is printed to stdout.

simple_local_anonymization.py

Detects personally identifiable information (PII) with the Piiranha model and replaces detected entities with [REDACTED].

Use as a library in Python:

from simple_local_anonymization import anonymize_text

with open("text_full.txt") as f:
    text = f.read()

clean = anonymize_text(text)
with open("text_anonymized.txt", "w") as f:
    f.write(clean)

Adjust the confidence threshold in detect_pii() if you get too many false positives or missed entities.

---

Audio file management

audio_renamer_simple.py

Sets ID3 tags (album_artist, artist, album) on all MP3 files under a given path. Edit my_tag and the root path (get_jpg_files("/K")) before running.

audio_renamer_multiple.py

Batch utilities for organizing audio libraries:

• rename_more_by_dir("./") — walks subdirectories and sets ID3 artist/album tags from each folder name
• find_duplicates(path) — removes duplicate MP3 files that share the same title tag
• convert_wma_to_mp3(input, output) — converts WMA files to MP3 via ffmpeg

The script runs rename_more_by_dir("./") on import. Comment out the last line or call functions manually as needed.

---

Ollama custom models

Shell scripts that create tuned Ollama models for coding tasks. Run once after pulling the base models.

ollm_qwen3-coder_modelfile.sh

Creates qwen3.5-coder from qwen3.5:9b with a 32k context window and coding-optimized parameters.

ollama pull qwen3.5:9b
bash ollm_qwen3-coder_modelfile.sh

ollm_qwen-pro_model_file.sh

Creates qwen-pro from qwen3-coder:14b-instruct-q4_K_M with a 32k context window, tuned for 16 GB RAM machines.

ollama pull qwen3-coder:14b-instruct-q4_K_M
bash ollm_qwen-pro_model_file.sh

Both scripts write a Modelfile in the current directory and call ollama create.

---

Quick start example

# 1. Set up environment
cp .env.example .env
# Edit .env and add HF_TOKEN if using speaker diarization

# 2. Convert audio
bash m4a_to_text.sh

# 3. Transcribe (pick one)
python neural_sound_to_text.py --audio_file Záznam.mp3
# python neural_sound_to_text_2_entities.py --audio_file Záznam.mp3 --num_speakers 2

# 4. Translate (requires Ollama)
ollama pull translategemma:12b
python simple_local_translator.py

---

Notes

• Memory: Whisper large-v3 and pyannote diarization are GPU/RAM intensive. On Apple Silicon, Whisper runs on MPS; diarization falls back to CPU for stability.
• Output file: Transcription scripts append to text_full.txt. Delete or rename the file before re-running if you want a fresh output.
• Secrets: Never commit .env. Only .env.example belongs in version control.