Working With Radio Lumieres Creole Content Pipeline
I spent about three years helping a small broadcasting collective in Port-au-Prince digitize and remaster their archive of community health and literacy broadcasts. The project involved taking dozens of hours of cassette recordings—many degraded by humidity, others recorded at inconsistent speeds—and converting them into a streamable format that could reach rural listeners via low-bandwidth mobile networks. It was messy, underfunded, and occasionally inspiring in ways that had nothing to do with technology. The core challenge wasn't technical so much as linguistic and cultural. Haitian Creole has regional variations that shift noticeably between the Artibonite Valley and the Sud department. A broadcast that resonates in one area can sound alien or even slightly offensive in another. Our team learned this the hard way when we attempted a uniform voice model for automated transcription, which produced results that were technically accurate but culturally tone-deaf. Some terms for agricultural advice translated differently depending on local farming practices. What we call "dry season planting" in one district might refer to something completely different elsewhere.
Transforming Haiti Radio Lumieres Lifechanging Broadcasts In Creole
Here is the practical breakdown of what went into making those broadcasts accessible, plus the actual workflow we used day to day. Audio restoration first. Before any transcription or translation work begins, the audio has to be cleaned. Most of our source tapes had background hiss, dropout issues, and occasional speed fluctuations from worn tape heads. We used iZotope RX 7 for noise reduction and spectral de-clicking. The noise floor on these recordings often sat around -42 dB, which is higher than you would see on a properly maintained archive. Standard denoise settings would wipe out the higher frequencies of Creole speech, making the content unintelligible. We settled on a manual subtraction approach, isolating the noise profile from silent sections and applying it selectively. This usually takes about 45 minutes per hour of source audio, but it preserves the full vocal range. Speed correction matters more than you think. Tape degradation causes wow and flutter—subtle speed variations that make voice recognition software fail. One particularly bad batch of tapes had a average pitch shift of about 3%. That sounds small until you run it through Whisper or Google's Speech-to-Text, and the output becomes garbage. We found that a two-pass approach works best. First, use Audacity's change pitch function with a targeted correction of minus 50 to minus 100 cents depending on the tape's condition. Then re-export and run through the ASR tool. This improved our transcription accuracy from roughly 62% to about 89% on the worst-condition source material.
Transcription in Haitian Creole is its own problem. The major speech-to-text models have limited support for Creole. Google's model is the best available as of my last check, but it still struggles with code-switching between Creole and French, which is extremely common in these broadcasts. Community health workers often use French medical terminology mid-sentence. Our workaround was building a custom grammar model on top of the base transcription. We created a phrase list of approximately 400 common medical and agricultural terms in both languages, then used it as a bias file during transcription. This pushed accuracy up another 8 to 12 percentage points on technical content. Without it, terms like "vaccination" or "malaria prevention" got mangled frequently. I remember one specific edge case that took us three days to solve. A particular broadcast had a speaker using a dialect from the Nord-Est region that included a lot of older Creole vocabulary not found in standard dictionaries. The ASR kept outputting nonsense phrases. We couldn't just ask the speaker to repeat it because the original recording was the only copy we had. What worked was pulling in a native speaker from that region, having them listen to the problematic segments, and manually transcribing those portions before stitching them back into the full file. It was slow work but necessary. Rushing it produced transcripts that were wrong in medically significant ways. Translation and localization. Once the transcription was clean, the next step was ensuring the language actually worked for the target audience. This wasn't about literal translation. It was about cultural adaptation. A health message about clean water might reference a well or a pump depending on the region. Our team worked with local community radio producers to review each script and flag anything that sounded off. This human-in-the-loop step typically added about 2 hours of work per hour of broadcast content, but it was non-negotiable for accuracy.
Get the Full Details
Encoding for low-bandwidth delivery. The final outputs needed to reach listeners on basic Android phones with limited data plans. We used Opus codec at 24 kbps mono, which cuts file sizes dramatically while keeping speech intelligible. A 30-minute broadcast comes out to roughly 20 MB instead of the 300+ MB you would get from an MP3 at standard bitrates. This was critical because many listeners in the target areas were paying per megabyte of data. The tradeoff is obvious—audio quality drops noticeably, but the alternative was simply not delivering the content at all. A workflow that actually held together. Here is the sequence we followed most of the time: Import source tapes into WaveLab or similar DAW. Run initial noise reduction pass. Check for speed irregularities using a reference tone if present, or estimate from speech patterns if not. Apply pitch correction. Export cleaned WAV. Run through Google ASR with the custom bias file. Review and correct the transcription manually, especially for technical terms. Send the corrected text to a native Creole speaker for localization review. Re-record or adjust the script as needed. Encode to Opus at target bitrate. Upload to the server and generate the stream. This pipeline usually took about 6 to 8 hours per hour of final content from start to finish with a small team of two to three people.
There are real bottlenecks in this process that nobody talks about. The biggest one is talent retention. The people who actually understand the linguistic nuances—the ones who can spot when a health term is being used incorrectly—are hard to find and even harder to pay consistently on nonprofit budgets. We had a great translator leave after eight months because the stipend didn't cover her actual cost of living in Port-au-Prince. You cannot automate that skill away. It requires lived experience with the language in a professional context. Another issue is source material scarcity. Many of the original broadcasts existed only on single copies of cassette tape that were physically deteriorating. We had about 40 hours of material in various states of decay, and roughly 15% of it was already lost to mold or binder hydrolyis before we could digitize it. That is a permanent loss. If you are working with an existing archive, prioritize the worst-condition tapes first and budget for potential total failure. For anyone attempting something similar, I would suggest starting small. Pick one hour of relatively clean source material and run it through the full pipeline before committing to a larger project. You will learn more from that one pass than from any amount of planning. The technical challenges are solvable. The cultural and linguistic ones require patience and genuine collaboration with native speakers, not just a quick translation pass.
The files we ultimately produced were distributed through a simple web player with an integrated download option. Listeners could stream at low bitrate or grab the Opus file for offline playback. We also provided the Creole transcripts as text files for use in community reading groups and by educators. The total archive grew to about 60 hours of processed content over the course of the project, though we never finished everything due to funding running out. If you are looking at this from a purely technical standpoint, the tools are all available for free or at low cost. The real difficulty lies in the human layer—the linguists, the cultural consultants, the people who understand that a broadcast about maternal health needs to sound like it comes from someone in the community, not a textbook. That is what makes or breaks this kind of work, and there is no shortcut around it.
