Skip to content

[BUG] Cohere Transcribe truncates audio (with fix) #736

Description

@eladve

Describe the bug

Summary

The "Cohere - High Accuracy" model in Fluid Voice silently only transcribes the beginning of the recording. The same model produces complete transcripts when run through HuggingFace's long-form chunking pipeline.

Root cause:
Cohere Transcribe's architecture requires chunked/long-form processing for audio longer than its feature extractor's max_audio_clip_s window. In single-pass mode, the model simply stops generating after one window's worth of audio.

The fix is in Cohere's own pipeline: the processor's audio_chunk_index output must be passed through to both model.generate() and processor.decode(). This triggers the chunked processing that reassembles per-chunk transcriptions. See: https://huggingface.co/CohereLabs/cohere-transcribe-03-2026 → "Long-form transcription" section.

Verification: running the same audio through the chunked pipeline locally produces the complete transcript with no truncation (tested with transformers 5.14.1, max_new_tokens=1024, audio_chunk_index passed through).

Environment

  • Fluid Voice 1.6.5, macOS
  • Model: "Cohere - High Accuracy" / Cohere Transcribe
  • Audio: 48-second mono recording, m4a format
  • macOS on Apple Silicon (M3 Pro)

Workaround / fix reference:

The fix is to use Cohere's long-form chunked pipeline, which passes audio_chunk_index through the generate/decode chain. Here's a minimal working example that produces complete transcripts:

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from transformers.audio_utils import load_audio

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-03-2026")
model = CohereAsrForConditionalGeneration.from_pretrained(
    "CohereLabs/cohere-transcribe-03-2026", device_map="auto"
)

audio_array = load_audio("speech.wav", sampling_rate=16000)

# Key: use audio= (numpy array), NOT raw waveform — this triggers chunking
inputs = processor(audio=audio_array, sampling_rate=16000, return_tensors="pt", language="en")
audio_chunk_index = inputs.get("audio_chunk_index")  # ← this is what's missing in single-pass
inputs.to(model.device, dtype=model.dtype)

outputs = model.generate(**inputs, max_new_tokens=1024)

# Pass audio_chunk_index to decode so chunks are reassembled
text = processor.decode(
    outputs, skip_special_tokens=True,
    audio_chunk_index=audio_chunk_index, language="en"
)
if isinstance(text, list):
    text = text[0]
print(text)

The CoreML port in Fluid Voice appears to run single-pass without chunking, which causes the model to stop generating after one feature-extractor window (~35s).

Reproduction steps

  1. In Fluid Voice -> Voice Engine, select "Cohere - High Accuracy (Cohere Transcribe)"
  2. Transcribe any audio file longer than ~35 seconds
  3. Compare the output to the same file transcribed via HuggingFace's pipeline

Expected behavior

Complete transcript of the entire audio file.

Actual behavior

The transcript cuts off after approximately 35 seconds of audio. No error or warning is shown — the truncation is silent, making it easy to miss that content is lost.

Example: a 48-second test recording consistently produces a transcript that stops mid-sentence at roughly the 35-second mark. The last ~13 seconds (containing 5 full sentences) are silently dropped. Tested twice with identical results.

App Version

Fluid Voice 1.6.5

macOS Version

MacOS Tahoe 26.5.1, Apple M3 Pro

Architecture

Apple Silicon

Logs or crash report


Screenshots or screen recording

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions